Content uploaded by Dayi Lin
Author content
All content in this area was uploaded by Dayi Lin on Aug 29, 2021
Content may be subject to copyright.
Towards a consistent interpretation of AIOps models
YINGZHE LYU∗,Software Analysis and Intelligence Lab (SAIL), Queen’s university, Canada
GOPI KRISHNAN RAJBAHADUR∗,Centre for Software Excellence, Huawei Canada, Canada
DAYI LIN∗,Centre for Software Excellence, Huawei Canada, Canada
BOYUAN CHEN∗,Centre for Software Excellence, Huawei Canada, Canada
ZHEN MING (JACK) JIANG, Lassonde School of Engineering, York University, Canada
Articial Intelligence for IT Operations (AIOps) has been adopted in organizations in various tasks, including
interpreting models to identify indicators of service failures. To avoid misleading practitioners, AIOps model
interpretations should be consistent (i.e., dierent AIOps models on the same task agree with one another
on feature importance). However, many AIOps studies violate established practices in the machine learning
community when deriving interpretations, such as interpreting models with suboptimal performance, though
the impact of such violations on the interpretation consistency has not been studied.
In this paper, we investigate the consistency of AIOps model interpretation along three dimensions: internal
consistency, external consistency, and time consistency. We conduct a case study on two AIOps tasks: predicting
Google cluster job failures, and Backblaze hard drive failures. We nd that the randomness from learners,
hyperparameter tuning, and data sampling should be controlled to generate consistent interpretations. AIOps
models with AUCs greater than 0.75 yield more consistent interpretation compared to low-performing models.
Finally, AIOps models that are constructed with the Sliding Window or Full History approaches have the
most consistent interpretation with the trends presented in the entire datasets. Our study provides valuable
guidelines for practitioners to derive consistent AIOps model interpretation.
CCS Concepts:
•Computing methodologies →Machine learning
;
•Software and its engineering →
Software maintenance tools;Maintaining software;Software evolution.
Additional Key Words and Phrases: AIOps, Model interpretation
ACM Reference Format:
Yingzhe Lyu, Gopi Krishnan Rajbahadur
∗
, Dayi Lin
∗
, Boyuan Chen
∗
, and Zhen Ming (Jack) Jiang. 2021. Towards
a consistent interpretation of AIOps models. ACM Trans. Softw. Eng. Methodol. , (August 2021), 39 pages.
https://doi.org/10.1145/nnnnnnn.nnnnnnn
1 INTRODUCTION
Ensuring the quality of service of cloud computing platforms is extremely pivotal. A recent IDG
survey [
32
] reports that 92% of organizations leverage cloud computing platforms for running their
applications. Therefore, it is extremely important to ensure that these cloud computing platforms
∗These authors contributed equally to the work.
Authors’ addresses: Yingzhe Lyu, ylyu@cs.queensu.ca, Software Analysis and Intelligence Lab (SAIL), Queen’s university,
Kingston, ON, Canada; Gopi Krishnan Rajbahadur
∗
, gopi.krishnan.rajbahadur1@huawei.com, Centre for Software Excellence,
Huawei Canada, ON, Canada; Dayi Lin
∗
, dayi.lin@huawei.com, Centre for Software Excellence, Huawei Canada, ON, Canada;
Boyuan Chen
∗
, boyuan.chen1@huawei.com, Centre for Software Excellence, Huawei Canada, ON, Canada; Zhen Ming
(Jack) Jiang, zmjiang@cse.yorku.ca, Lassonde School of Engineering, York University, ON, Canada.
Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee
provided that copies are not made or distributed for prot or commercial advantage and that copies bear this notice and
the full citation on the rst page. Copyrights for components of this work owned by others than ACM must be honored.
Abstracting with credit is permitted. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires
prior specic permission and/or a fee. Request permissions from permissions@acm.org.
©2021 Association for Computing Machinery.
1049-331X/2021/8-ART $15.00
https://doi.org/10.1145/nnnnnnn.nnnnnnn
ACM Trans. Softw. Eng. Methodol., Vol. , No. , Article . Publication date: August 2021.
2 Yingzhe Lyu, Gopi Krishnan Rajbahadur∗, Dayi Lin∗, Boyuan Chen∗, and Zhen Ming (Jack) Jiang
remain highly available and ecient, particularly since failures in cloud computing platforms are
estimated to cost up to $700 billion annually [
58
]. For instance, a recent survey [
4
] pegs the average
cost per hour of an organization’s server downtime to be anywhere between $300,001 to $400,000.
Cloud computing platforms generate a tremendous amount of data which is impossible to be
analyzed manually. Recently, it has become increasingly common for organizations to use AIOps
(Articial Intelligence for IT Operations) to leverage such generated data to ensure the quality of
service and high availability of cloud computing platforms [
6
,
8
,
21
,
42
,
62
,
88
]. AIOps leverages
machine learning learners to construct machine learning models (hereafter AIOps models) with
operations data collected from the cloud computing platforms (e.g., logs and alert signals) to enable
quality assurance tasks such as predicting hard drive failures [
42
], job termination [
22
], service
outages [
88
] and performance issues [
44
]. Note that we use the term “learner” to refer to a machine
learning algorithm (e.g., Random Forest) and the term “model” to refer to a trained machine learning
model (e.g., a Random Forest model trained on disk failure data).
In addition to using AIOps models to predict failures and outages on a cloud computing platform,
several prior studies also interpret AIOps models to identify association between dierent factors
and occurrences of certain failures or outages to make operational and business decisions. For
instance, Chen et al
. [17]
interpret their deep learning based incident prediction model and nd that
incident-occurring environment features are one of the most important indicators of a potential
incident happening in their platform. Similarly, Zhao et al
. [88]
interpret their XGBoost based
incident prediction model and nd that in their systems, database related issues are the root
cause for the incidents. Li et al
. [42]
interpret their AIOps models to rene their automated alert
management system.
These derived interpretations of AIOps models should be consistent (i.e., the feature importance
ranking from dierent models’ interpretations agree with one another), to avoid misleading practi-
tioners. For instance, as we mentioned earlier, Chen et al
. [17]
derive interpretations from their deep
learning based incident prediction model. However, if their incident prediction model is retrained
(e.g., on a local instance of the same training data) and it produces a dierent set of features being
the most important feature associated with incidents, the practitioners might not know which
interpretation to trust. Such a case is indeed possible as Pham et al
. [64]
show, when a model
is retrained, even on the same training dataset, its performance might change (sometimes even
drastically). Thus, it is possible that the interpretations that are derived from the retrained incident
prediction model might also vary. More generally, recent studies [
75
,
76
] in several domains point
out that many factors (e.g., dierent hyperparameters) could impact the consistency of derived
interpretations of a machine learning model.
However, to the best of our knowledge, none of the prior studies in the eld of AIOps investigates
the factors that could impact the consistency of the derived interpretations of AIOps models.
Though the factors relating to the consistency of derived interpretations have been explored by a
handful of studies [
36
,
66
,
75
,
76
] in other domains, it is pivotal to explore them in the context of
AIOps for the following reasons:
•
Chen et al
. [20]
and Dang et al
. [21]
point out there are several challenges that are unique to
the eld of AIOps. These unique challenges might present unique factors (which are typically
not prevalent in other domains such as machine learning) that impact the consistency of
the interpretations derived from AIOps models dierently. For instance, as Dang et al
. [21]
explain, AIOps models are constantly updated/retrained to keep up with the rapidly evolving
data. Such rapid retraining of AIOps models could mean that the derived interpretations of an
AIOps model could change with every retraining. However, as we explain earlier, consistency
of derived interpretation is pivotal for practitioners. Therefore, it is important to understand
ACM Trans. Softw. Eng. Methodol., Vol. , No. , Article . Publication date: August 2021.
Towards a consistent interpretation of AIOps models 3
how the factors that are unique to AIOps impact the consistency of interpretations derived
from AIOps.
•
The operations data that is typically used to build AIOps models is very dierent from the
data used in other domains like machine learning. Operations data is typically a mixture of
temporal (e.g., log data) and spatial (e.g., hardware congurations) data. In addition, it may
also contain a mixture of heterogeneous data types like numeric, ordinal and nominal values,
making it starkly dierent from data used in other machine learning tasks (which is typically
either spatial or temporal in nature).
•
As Menzies
[57]
and Ray et al
. [67]
show, ndings from other domains like machine learn-
ing do not necessarily generalize in software analytics domains like AIOps. For example,
both Menzies
[57]
and Ray et al
. [67]
show that language models used in the machine
learning domain typically yield spurious results on software engineering related data. Fur-
thermore, Menzies
[57]
warns us that simply using the methods and ndings outlined in the
eld of machine learning on software engineering data might yield suboptimal results.
Therefore, in our study, we aim to better understand the factors that could impact the consistency
and the practical adoption of the derived interpretations of an AIOps model. In our paper, we
rst propose a set of rigorous criteria to thoroughly assess the consistency of the interpretations
derived from AIOps models. We do so through a case study on two publicly available operations
datasets (the Google cluster trace dataset [
68
] and the BackBlaze hard drive statistics dataset [
5
])
that have been widely used by many prior studies in AIOps [
11
,
22
,
71
]. In particular, our proposed
criteria investigate how dierent factors impact the AIOps model interpretation along three key
dimensions: Internal consistency,External consistency, and Time consistency.
•Internal consistency
[
64
] (a.k.a., model reproducibility) captures the similarity between
the derived interpretations of an AIOps model trained from the same setup (i.e., same train-
ing data and same implementation) across multiple executions. In other words, internal
consistency checks if the interpretations derived from an AIOps model is reproducible. For
instance, as Pham et al
. [64]
show, and a plethora of prior studies [
26
,
31
,
64
,
65
] warn, unless
randomness involved during the training process of machine learning models is controlled,
the results might not be reproducible. As a result, the derived interpretations might not be
internally consistent which would make it impossible for the managers and DevOps engineers
to choose which interpretation to act on. Despite that, many of the AIOps studies do not take
specic steps to ensure reproducibility. For instance, both Zhao et al
. [88]
and Li et al
. [42]
use deep learning models to construct their AIOps models. However, neither of those studies
take explicit methods to control the randomness involved which could in turn, impact the
reproducibility of the interpretation derived from these AIOps models. Therefore in RQ1
(Are the interpretations from AIOps models internally consistent?), we study the impact of
potential sources of randomness during the model training, which can impact the internal
consistency of the derived interpretations of the AIOps models.
Results.
All the three studied sources of randomness (i.e., inherent randomness from learners,
randomized hyperparameter tuning and data sampling randomness) impact the consistency
of AIOps model interpretation. Sampling randomness introduces the largest scale of incon-
sistency to AIOps model interpretation. When all the three sources are controlled during
the training of an AIOps model, the derived interpretation of an AIOps model is internally
consistent.
•External consistency
[
66
] captures the similarity between the derived interpretations of
similar-performing AIOps models on a given dataset. In other words, external consistency
sanity checks if the models that have the similar performance report similar interpretations
ACM Trans. Softw. Eng. Methodol., Vol. , No. , Article . Publication date: August 2021.
4 Yingzhe Lyu, Gopi Krishnan Rajbahadur∗, Dayi Lin∗, Boyuan Chen∗, and Zhen Ming (Jack) Jiang
for a dataset. Typically, interpretable models are preferred by the practitioners to derive
interpretations. Several recent studies in AIOps [
42
] and machine learning [
72
] argue that
interpretable models even with lower performance are preferred in the AIOps context when
high-performing models are not easily interpretable. Rajbahadur et al
. [66]
recently show that
the derived interpretation between dierent machine learning models could vary considerably.
In addition, Lipton
[49]
and Molnar
[61]
warn that such inconsistency could be exacerbated if
low-performing models are used to derive interpretations. Intuitively, interpretations derived
from a low-performing interpretable model could be trustworthy only if the interpretable
model has the same interpretation as other machine learning models on a given dataset (at
least among the similar-performing models). In other words, dierent models at the same
performance level should generate similar interpretations. However, there are no studies
which examine the relation between the model performance and model interpretations.
In particular, it is not clear if using high-performing models would yield more consistent
interpretations than using low-performing models. Hence, in this paper, we assess and
compare the external consistency of the interpretations from models at dierent performance
levels (i.e., similarities of model interpretations among low-performing models vs. similarities
of model interpretations among high-performing models). Such study is even more important
for low-performing models since several prior studies question if low-performing models
might generate inconsistent interpretations [
37
,
49
,
61
]. Otherwise, interchangeably using
models to derive interpretations might be misleading and might result in misguided decisions.
Therefore in RQ2 (Are the interpretations from AIOps models externally consistent?), we
investigate the external consistency of derived interpretations of AIOps models at dierent
performance levels.
Results.
The interpretations derived from high-performing AIOps models (with a minimum
acceptable AUC of 0.75) are more consistent than those from low-performing AIOps models.
The interpretations derived from high-performing AIOps models exhibit strong external
consistency.
•Time consistency
[
42
] captures the similarity between the derived interpretations of an
AIOps model across dierent time periods (i.e., as the AIOps models evolve). In other words,
time consistency checks if the interpretation derived from an AIOps model remains general-
izable across time. Interpretations derived from AIOps models may be used to make business
decisions and process optimizations that have a long-lasting eect. Hence, it is pivotal that
the interpretation of an AIOps model should not only just reect the trend from the most
recent data on which it was trained, but also should capture and reect the trends observed
over a longer period of time. However, previous works in defect prediction [
7
] and AIOps [
42
]
show machine learning models trained on one time period do not generalize well when tested
on a dierent time period and their derived interpretations could also vary depending on
the size of the training data. Therefore, in RQ3 (Are the interpretations from AIOps models
consistent across time (i.e., time consistent)?), we examine which of the commonly used
model updating approaches allows the interpretations of the updated models to be consistent
over long periods of time.
Results.
The derived interpretations from AIOps models constructed with the Sliding Win-
dow and Full History approaches best capture the trends present across time periods in the
entire dataset and exhibit strong time consistency.
The main contributions of our paper are the following:
(1)
This is the rst work that studies the factors that impact the consistency of AIOps model
interpretations.
ACM Trans. Softw. Eng. Methodol., Vol. , No. , Article . Publication date: August 2021.
Towards a consistent interpretation of AIOps models 5
(2)
We propose a set of rigorous criteria that enables researchers and practitioners to assess the
consistency of their derived interpretations from AIOps models.
(3)
We provide several actionable guidelines to improve the consistency of interpretations derived
from AIOps models.
(4)
To foster replicability of our ndings and promote open science, we make the datasets and
the code to conduct our study publicly available.
Paper organization. The rest of the paper is organized as follows: in Section 2 we present a mo-
tivational example that motivates the need for our study in AIOps. In Section 3 we provide the
background and related work of AIOps, and introduce the research questions that we investigate in
our paper. In Section 4 we explain our case study setup. Section 5, Section 6, and Section 7 present
each of our RQs. Section 8 discusses the results, limitations of the results and potential future areas
for investigation. In Section 9, based on our ndings from the results of the studied RQs, we outline
several practical guidelines for AIOps researchers and practitioners. In Section 10 we discuss the
threats to validity of our study. Finally, in Section 11 we conclude our study.
2 A MOTIVATIONAL EXAMPLE
Lei is a DevOps engineer. They are responsible for monitoring and maintaining a batch processing
job that runs daily. The batch processing job runs on a cluster and can take hours to nish. In the
past, they have noticed that the job may randomly fail and they would need to re-submit the job
for execution. Recently the analytics team helped Lei build an AIOps model, which uses the traces
collected from the cluster to predict if the job is going to fail in the next half hour. This model is
selected from a pool of best-performing candidate models by the analytics team after considering
various aspects such as performance and training costs. The analytics team advised Lei to use the
AIOps model in two ways:
(1) Predictive
: When the model predicts the job is going to fail in the next half hour with high
condence, they can manually terminate and re-submit the job, to save execution time.
(2) Explanatory
: By interpreting the trained model, the analytics team advised them that the
model learnt that one of the most impactful factors that is associated with the job failure
is the launch of a specic competing job from another team. Therefore from an operations
perspective they should contact the other team to schedule the competing job at a dierent
time of the day.
Scenario 1
: To productionize the model, the analytics team provided Lei with the data and code
used to train the model. Lei executed the provided training code on the same data, only to realize
that the model they trained provided a dierent interpretation compared to that of the model
trained by the analytics team. In particular, the model they trained does not consider the competing
job as an important factor. Lei is confused about which interpretation to trust, and whether they
should reschedule the competing job with the other team. In this scenario, Lei is concerned about
the internal consistency of the AIOps model.
Scenario 2
: To avoid performance drift of the model in production, the model needs to be
periodically retrained with latest data. However the current model provided by the analytics team
is too costly to retrain. Lei asked the analytics team if there is any alternative candidate model that
is cheaper to retrain. The analytics team sent over a new, lighter model, which is cheaper to retrain
albeit at a slightly reduced performance. However, Lei realized that even though the new model
was trained on the same data as the current one, it provides a dierent interpretation, and does
not consider the competing job as an important factor, posing a dilemma for Lei about whether
to reschedule the competing job with the other team. In this scenario, Lei is concerned about the
external consistency of the AIOps model.
ACM Trans. Softw. Eng. Methodol., Vol. , No. , Article . Publication date: August 2021.
6 Yingzhe Lyu, Gopi Krishnan Rajbahadur∗, Dayi Lin∗, Boyuan Chen∗, and Zhen Ming (Jack) Jiang
Scenario 3
: The model was deployed in production and was scheduled to be retrained with new
data everyday at midnight. Lei contacted the other team about rescheduling the competing job, as
advised by the analytics team based on the interpretation of the deployed model. Rescheduling jobs
may lead to a lot of downstream administrative and operative changes, and therefore is expensive
to communicate and perform. However, in just a few days, before the rescheduling was scheduled
to take eect, Lei noticed that the recently updated model no longer considers the competing job
as an important factor. Such drastic changes in the model interpretation led to a lot of confusion
among teams, and wasted eort in rescheduling jobs. In this scenario, Lei is concerned about the
time consistency of the AIOps model.
3 BACKGROUND
In this section, we rst describe the existing work on AIOps. Then we present the current practices
of deriving AIOps interpretation. Finally, we present our criteria for assessing the consistency of
AIOps interpretation.
3.1 Existing work on AIOps
We rst introduce the common AIOps applications. Then we discuss the existing work on AIOps
interpretation. Finally, we present the reproducibility concerns in general machine learning tasks.
3.1.1 AIOps Applications. Although huge eorts have been devoted to cloud computing systems
for ensuring the quality of services, various types of issues (e.g., job termination, hard drive failure,
and performance anomalies) are unavoidable. To ensure the reliability of online services, we must
resolve and manage these issues in a timely manner, as failing to do so might cause unavailable
services and huge nancial losses. AIOps solutions contribute to issue management in two phases:
(1) AIOps solutions aim to predict whether certain issues would occur by learning from the historical
data; and (2) after the issues occur, AIOps solutions aim to help mitigate the issues (e.g., automated
problem diagnosis) or provide suggestions to the domain experts (e.g., incident triage).
Issue Prediction.
Many of the prior works focus on analyzing monitoring data for predicting the
occurrence of various types of issues [11, 19, 22, 29, 40–42, 44, 47, 56, 71, 82, 88].
Lin et al. [
47
] and Li et al. [
42
] predict node failures in large-scale cloud computing platforms
by building machine learning models from temporal (e.g., CPU utilization metrics), spatial (e.g.,
location of a node), and cong data (e.g., build data). Similarly, Li et al [
40
,
41
] build tree-based
models to predict hard drive failures. Botezaku et al. [
11
], Mahdisoltani et al. [
56
], and Xu et al. [
82
]
leverage SMART-based analysis to build a machine learning pipeline for predicting hard drive
failures in large-scale cloud computing platforms. Chen et al. [
19
] collect and analyze the alert data
and its dependencies to predict outages in the whole cloud systems. Zhao et al.[
88
] propose a deep
learning-based approach, eWarn, which leverages textual (e.g., keywords in incident tickets) and
statistical features (e.g, alert count) to predict incident occurrences. El-Sayed et al. [
22
] and Rosa
et al. [
71
] predict job failures from trace data collected from Google cloud computing platform.
Lim et al. [
44
] leverage performance metrics to cluster performance issues into the recurrent and
unknown ones.
Issue Mitigation.
Issues in online services need to be mitigated in a timely manner. Many of the
prior works focus on triaging [
8
,
16
,
17
], diagnosing [
6
,
18
,
33
,
53
,
87
], and managing issues [
35
,
48
,
50, 51, 83, 84], which benet the mitigation process.
Triaging. Chen et al. [
16
] propose a deep learning-based technique to improve the current incident
triage process (e.g., distributing the new incident to the responsible team). Chen et al. [
17
] perform
an empirical study on characterizing incidents in online systems and propose DeepIP, a technique
to detect incidental incidents (i.e., incidents that are less severe and last for a short period of time),
ACM Trans. Softw. Eng. Methodol., Vol. , No. , Article . Publication date: August 2021.
Towards a consistent interpretation of AIOps models 7
which can reduce the incident triage eorts. Bansal et al. [
8
] propose DeCaf, a Random Forest-based
framework to correlate telemetry data with performance regressions. In addition, the detected
performance regressions are automatically triaged to on-site engineering team.
Diagnosing. Zhang et al. [
87
] propose an ensemble of models to automatically diagnose perfor-
mance problems. Chen et al. [
18
] propose LiDAR, a deep learning-based approach to linking similar
incidents based on historical information. Luo et al. [
53
] mine time-series data and event data to
discover correlations between them, which could improve the incident diagnosis process. Baner-
jee et al. [
6
] discuss challenges in performance diagnosis in a hybrid-cloud enterprise software
environment. Jehangiri et al. [
33
] present techniques to diagnose performance anomalies using
time-series datasets.
Managing. Jiang et al. [
35
] analyze the similarity between incident descriptions and their corre-
sponding troubleshooting guide to facilitate incident management. Lou et al. [
50
,
51
] develop a
software analytic-based system to resolve scalability, reliability, and maintainability of data-driven
incident management systems. Xue et al. [
83
,
84
] proactively reduce performance tickets by pre-
dicting usage series in cloud data centers. Lin et al. [
48
] propose a data mining-based technique to
detect emerging issues (a sudden burst of new issues) by analyzing historical issues.
3.1.2 AIOps Interpretation. The importance of interpretability of machine learning-based systems
has drawn increasing attention recently as it is related to the trustworthiness, reliability, and quality
of the systems. In the context of AIOps, it is essential that we can reason about the model recom-
mendations to make business decisions (e.g., replacing failing hard drives) and improve the status
quo (e.g., improving monitoring infrastructure). Here we discuss prior studies on interpretations of
AIOps solutions.
Li et al. [
42
] study the importance scores of a random forest-based model and nd that alert data is
the most important feature set for node failure prediction. Zhao et al. [
88
] leverage a model-agnostic
technique, LIME [
69
], to generate interpretable report for incident prediction. This technique is
mainly used for interpreting each individual prediction. Bansal et al. [
8
] propose a technique to
extract ranked list of rules from random forest-based models, which can be used to explain predicted
performance regression. Li et al. [
40
,
41
] present that by using interpretable models (e.g., regression
tree-based models), users can derive more meaningful decisions in reducing the hard drive failure
rate, which is preferred over ANN-based models in such context. Chen et al. [
16
] and Jiang et
al. [
35
] all leverage deep learning-based systems for mitigating incidents. The deep learning-based
systems contain multiple components for reducing the data noise and adjusting loss functions.
They interpret the importance of these components by comparing the performance before and
after enabling these services.
3.1.3 Reproducibility of Machine learning. Recent studies [
26
,
31
] have shown that reproducibility
has drawn increasing attention in the machine learning eld in recent years. Many research
papers did not include sucient information (e.g., dataset and experiment code), which makes it
dicult for other researchers to reproduce the experiment results. Furthermore, even with the same
experiment setup, there is still variance in the outcomes due to the stochastic nature of machine
learning applications [
64
]. Hence, huge eorts have been devoted to evaluating and improving the
reproducibility in the machine learning eld. Pham et al. [
64
] study the variance of deep learning
applications and nd that non-implementation level factors can cause the accuracy and training
time to uctuate. Pineau et al. [
65
] propose a reproducibility checklist before submitting papers to
NeurIPS, to improve the quality of scientic contributions. Various guidelines (e.g., Model Cards [
60
]
and Datasheets [
25
]) have also been proposed to help practitioners and researchers to improve the
reproducibility of their machine learning applications.
ACM Trans. Softw. Eng. Methodol., Vol. , No. , Article . Publication date: August 2021.
8 Yingzhe Lyu, Gopi Krishnan Rajbahadur∗, Dayi Lin∗, Boyuan Chen∗, and Zhen Ming (Jack) Jiang
Preprocessing
Monitoring
data
Feature
engineering
Training
dataset
Model training
Hyper-
parameter
tuning
Data
sampling
ML
learners
Model evaluation Model evolution
Trained
models Evaluated
models Updated
models
Testing
dataset
Performance
evaluation
Model
update
strategy
Model
update
New
monitoring
data
Model interpretation RQ1:
Internal consistency RQ2
External consistency RQ3
Time consistency
Fig. 1. An overall process of deriving interpretations from AIOps models, our criteria for assessing the
consistency of AIOps model interpretation is highlighted along the process.
The interpretation of machine learning applications also suers from the reproducibility issues.
Fan et al. [
23
] assess the quality of ve interpretation techniques in Android malware analysis
applications, where they evaluate the stability, robustness, and eectivenss of the interpretations.
Warnecke et al. [
81
] also study the similar dimensions of interpretation in security domain. They
both nd that dierent interpretation techniques could generate dierent interpretation results
for the same prediction result. Other studies [
3
,
14
,
85
] also indicate that the interpretation results
cannot be trusted when the results cannot be reproduced.
Our study is dierent from the previous work in three aspects: (1) our study focuses on the consis-
tency of AIOps interpretation, in particular, we study the internal, external, and time consistency by
leveraging one interpretation technique; (2) we leverage the model-level interpretation technique,
i.e. the feature importance ranking, to extract the general trend and knowledge in the AIOps model,
where previous studies mostly focus on instance-level interpretation; and (3) our work is conducted
in the AIOps domain with constantly evolving data, which requires the interpretation to be able to
capture the general trends across time.
3.2 The current practices of deriving interpretations from AIOps models
We rst introduce the the overall process of deriving interpretations from AIOps models. Then we
discuss the potential issues with recent AIOps studies.
3.2.1 The overall process. As we discussed in the previous section, AIOps solutions have been
used extensively for a variety of tasks (e.g., incident prediction [
42
,
88
], performance analysis [
6
,
8
],
anomaly detection [
62
], and business decision making [
21
]). The quality of AIOps solutions is
evaluated from various dimensions such as performance, scalability, and maintainability. Although
many studies leverage the interpretations of AIOps models to support critical decisions, few studies
have been done to systematically assess the quality of interpretations of AIOps models.
Figure 1 shows an overall process of deriving interpretations from AIOps models. We divide the
pipeline into four major phases:
(1)
The Preprocessing phase refers to the process of continuously collecting monitoring data and
transforming it into readily available features as inputs for ML models. This is a common
practice adopted in many AIOps solutions [
42
,
55
]. In this phase, the training dataset is
generated.
ACM Trans. Softw. Eng. Methodol., Vol. , No. , Article . Publication date: August 2021.
Towards a consistent interpretation of AIOps models 9
Table 1. The potential neglected key dimensions of interpretation consistency.
Paper Internal Consistency External Consistency Time Consistency
P1 [42] # # #
P2 [88] #
P3 [8] # #
P4 [40] #
P5 [41] # #
P6 [30] # #
P7 [86] # #
P8 [17] #
P9 [35] # #
P10 [16] # #
P11 [47] # #
(2)
The Model Training phase refers to the process of training ML models using preprocessed
features. It consists of three steps: data sampling, ML learner tting, and hyperparameter
tuning. Data sampling is for selecting a representative subset of a large dataset. ML learner
tting is to select the appropriate learner for the task and t a model with the learner.
Hyperparameter tuning is to nd the optimal parameters for the models. In this phase, the
trained models are generated.
(3)
The Model Evaluation phase refers to evaluating the performance of the trained models using
the testing dataset. The testing dataset is not seen by the model during the model training
phase to prevent data leakage. In this phase, the evaluated models are generated.
(4)
The Model Evolution phase refers to the practice of deploying the models and constantly
updating models. Various model update strategies are applied to reect the trends contained in
the newly available period of monitoring data and mitigate the impact of concept drift [
24
,
28
].
In this phase, the updated models are generated.
The overall process is iteratively conducted until the model performance is above a certain
threshold. The interpretation can be derived from the trained models, evaluated models, and the
updated models. There are two general approaches to interpret machine learning models: the
model-specic approach and the model-agnostic approach. The model-specic approach is mainly
used with models that are intrinsically interpretable. Examples of intrinsically interpretable models
include linear regression models and decision tree-based models. On the other hand, some machine
learning models are dicult to interpret due to their complex internal structures (e.g., deep neural
networks). To interpret these models, practitioners mainly use the model-agnostic approach which
explains the models in a post-hoc manner. There are many model-agnostic techniques in the
existing literature, such as LIME [
69
] and permutation feature importance. There are two types
of interpretation results that could be produced: model-level interpretation and instance-level
interpretation. The model-level interpretation is to understand how the models work internally,
while the instance-level interpretation is to explain each single prediction. In this paper, we use the
permutation feature importance, which is a model-level, model-agnostic interpretation approach, to
interpret all the studied AIOps models. We explain the approach and rationale in detail in Section 4.
ACM Trans. Softw. Eng. Methodol., Vol. , No. , Article . Publication date: August 2021.
10 Yingzhe Lyu, Gopi Krishnan Rajbahadur∗, Dayi Lin∗, Boyuan Chen∗, and Zhen Ming (Jack) Jiang
3.2.2 Potential issues with recent AIOps studies. In order to understand whether the recent AIOps
studies account for the three aforementioned key dimensions along the interpretation consistency,
we conducted a literature survey of AIOps studies that derive interpretations from their AIOps mod-
els. To survey the literature, we searched Google Scholar with terms including "AIOps", "Software
Engineering", "Hard Drive Failures", and "Incident Prediction" to collect the initial set of studies.
We then ltered these studies to keep the ones that were published in the last seven years (i.e., after
2014). In the end, eleven studies that we included in our survey were carefully examined by one of
the authors. The results are shown in Table 1.
According to our survey, none of the surveyed papers explicitly mentioned that they control
the randomness from the learners, data sampling, and hyperparameter tuning, where applicable. 2
out of 11 studies suggested interpreting low-performing models to derive interpretations. While
7 out of 11 studies did not consider or discuss the impact of periodic retraining strategies on the
interpretation consistency.
Our summary by no means intends to criticize the prior studies, but to raise awareness that
there is currently a gap between recent AIOps studies and the commonly followed practices (e.g.,
controlling the randomness) in the machine learning community. It is important to take into account
if the interpretations derived from AIOps models are internally consistent, externally consistent, or
time consistent to avoid misleading decisions from being made by practitioners.
3.3 Our criteria for assessing the consistency of AIOps interpretation
Here we describe the motivation of assessing each dimension of AIOps model interpretation
consistency in details and formulate the corresponding RQs.
Internal Consistency.
Prior work has shown the impact of dierent random factors (e.g., random
sampling of the same dataset, and dierent random seeds for tting ML learner) in the reproducibility
of predictive software engineering [
43
] and deep learning [
64
] research, with the focus on model
performance reproducibility. As a result, the derived interpretations might not be reproducible
(a.k.a internally consistent). Despite that, many of the AIOps studies did not take actions to ensure
reproducibility. To investigate how the randomness from the three steps in the model training
phase (highlighted in Figure 1) impacts the internal consistency of interpretations derived from
trained models, we formulate the following research question:
(RQ1) Are the interpretations from AIOps models internally consistent?
External Consistency.
Many existing studies use interpretable models to compute feature impor-
tance ranks without taking into consideration the performance of these interpretable models. The
assumption is that these interpretable models are able to capture the general trends in the data
despite their lower performance. However, several studies already hint that such a practice could
cause inconsistency in the computed insights [
49
,
61
,
66
]. Hence, in the model evaluation phase
(highlighted in Figure 1), to assess the consistency of the interpretations derived from the evaluated
models, we formulate the following research question:
(RQ2) Are the interpretations from AIOps models externally consistent?
Time Consistency.
The workloads of large cloud computing platforms are highly dynamic and
they tend to evolve signicantly throughout its lifetime [
42
,
47
]. To keep up with the constantly
evolving, temporal nature of the data, prior studies typically use several model update approaches to
keep their AIOps models current and relevant. For instance, some studies periodically retrain their
ACM Trans. Softw. Eng. Methodol., Vol. , No. , Article . Publication date: August 2021.
Towards a consistent interpretation of AIOps models 11
AIOps models on new data [
42
,
47
] or construct time-based ensemble approaches that aggregate
local AIOps models trained on small time periods [
73
,
80
]. These models are typically interpreted to
make operational decisions [
42
,
72
]. However, the temporal nature of the data might make it hard
for the dierent model update approaches to update the AIOps models to accurately generalize
to the underlying trends while still yielding high performance on the most recent data. Hence, to
avoid misleading operational decisions, it is important to prevent these updated AIOps models
from losing sight of the historical trends in data and overtting to the most recent time period.
To examine which of the commonly used model updating approaches allows the interpretations
of updated models to have the highest time consistency in the model evolution phase (highlighted
in Figure 1), we formulate the following research question:
(RQ3) Are the interpretations from AIOps models consistent across time (i.e., time consistent)?
4 CASE STUDY SETUP
In this section, we describe the setup of our case study in details.
4.1 Studied AIOps Datasets
To understand the challenges of reliably interpreting AIOps models, we perform a case study on two
large-scale public AIOps datasets that have been commonly used in prior work [
11
,
22
,
55
,
56
,
71
,
82
]:
the Google cluster trace dataset [
68
] (hereinafter referred to as the Google dataset), and the Backblaze
hard drive statistics dataset [5] (hereinafter referred to as the Backblaze dataset).
4.1.1 The Google dataset. The Google dataset contains the trace information of job runs on a
large-scale cluster at Google. In this study, we use the second version of the dataset
1
that was
collected on May 2011, containing 29 days of trace information from a cluster of about 12.5K
machines. The dataset includes information about the machines in the cluster, the jobs executed on
the cluster, and the tasks under each job. In total, there are 670K jobs and 26M tasks in the dataset.
There are four possible states of a job in its lifecycle: unsubmitted, pending, running, and dead.
Throughout the lifecycle, a job triggers multiple events that are recorded in the dataset, including
submit, schedule, evict, fail, kill, nish, lost, and update.
4.1.2 The Backblaze dataset. The Backblaze dataset contains daily snapshot of statistics of the hard
drives in the Backblaze data center. The Backblaze dataset includes hard drive information (e.g.,
the model and capacity of the hard drive) and SMART (Self-Monitoring, Analysis and Reporting
Technology) metrics of the hard drives. SMART is a hard drive monitoring system that monitors
various indicators of drive reliability, to identify imminent hard drive failures. In this study, we use
36 months of Backblaze snapshot data collected from 2015 to 2017, containing 72M records.
4.2 Experiment Context
In this case study, we focus on two AIOps applications: cluster job failure prediction on the Google
dataset, and hard drive failure prediction on the Backblaze dataset. The interpretation of cluster
job failure prediction models can help practitioners identify factors that increase the likelihood
of job failure, apply fail-safe mechanisms at the time of job submission, and investigate methods
to reduce the occurrence of such risky factors. The interpretation of hard drive failure prediction
models can help practitioners understand the early indicators of hard drive failures and improve
monitoring infrastructure around the early indicators. Below we describe the general experiment
1https://github.com/google/cluster-data/blob/master/ClusterData2011_2.md
ACM Trans. Softw. Eng. Methodol., Vol. , No. , Article . Publication date: August 2021.
12 Yingzhe Lyu, Gopi Krishnan Rajbahadur∗, Dayi Lin∗, Boyuan Chen∗, and Zhen Ming (Jack) Jiang
Table 2. Description of metrics for Google cluster job failure prediction.
Metric Description
Scheduling class A number that aects policies for resource access.
Num Tasks The number of tasks in a job.
Priority The priority of the job.
Dierent machine
Whether a task must be scheduled to execute on a dierent machine than any
other currently running task in the job.
Requested CPU Requested CPU resources.
Requested Disk Requested disk space resources.
Mean CPU usage Mean CPU usage over 5 minutes after job submission.
Mean Memory usage Mean memory usage over 5 minutes after job submission.
Mean Disk Usage Mean disk usage over 5 minutes after job submission.
Sd Memory usage Standard variation of the memory usage over 5 minutes after job submission.
context of our case study. We describe the RQ-specic experiment designs in the sections of each
RQ. A replication package of our experiments and analysis are also provided 2.
4.2.1 Data preprocessing. Here we discuss the data preprocessing steps for two datasets separately.
Google cluster job failure prediction:
To predict job failures on the Google dataset, we need
to rst identify whether a job in the dataset nished successfully or not. Such information is not
directly provided in the dataset. We use the events of jobs in the dataset as indications for such
information. Specically, we consider a job as failed if its last event in the dataset is a “fail” event.
To avoid mislabelling jobs that have not nished running at the time of data collection, we exclude
the jobs that started on the last day available in the dataset.
Similar to prior work [
22
,
55
], we extract a set of temporal and conguration metrics as candidate
features for training the predictive models. Tantithamthavorn and Hassan [
74
] stated that highly-
correlated features should not be used to construct models for interpretation. Therefore, to remove
the correlated and redundant features, we use the
varclus
function and the
redun
function in
the
Hmisc
R package to identify and remove metrics that show high Spearman correlation and
multicollinearity. Table 2 provides a description of metrics used in our study for Google cluster job
failure prediction.
We then divide the dataset into subsets with equal time intervals (i.e., periods), based on the date
when the job is created. In total, we divide the dataset into 28 one-day periods.
Backblaze hard drive failure prediction:
To predict hard drive failures on the Backblaze
dataset, we extract the same set of metrics as prior work [
55
] as candidate features. The dataset is
divided into subsets of one-month intervals (i.e., periods), to allow us label the hard drives that fail
in the next period and associate them with their metrics in the current period. In total, we obtained
36 one-month periods, with the metrics of each hard drive in the period, and whether each hard
drive fails in the next period.
Similar to the Google dataset, we remove metrics that show high Spearman correlation and
multicollinearity. Table 3 provides a description of metrics used in our study for Backblaze hard
drive failure prediction.
4.2.2 Model training. To ensure the result of our case study is generalizable, we include a variety
of learners in our study that have been used in literature [
11
,
22
,
55
,
56
]: Linear Discriminant
Analysis (LDA), Quadratic Discriminant Analysis (QDA), Logistic Regression (LR), Classication
2
https://github.com/YingzheLyu/AIOpsInterpretation. The replication package is set to be private during the review period,
and will be made public once the paper is accepted. Please use the following GitHub account to view the repository in the
meanwhile. Username: reviewer-AIOpsInterpretation, Password: gUU!9sL9UDPT
ACM Trans. Softw. Eng. Methodol., Vol. , No. , Article . Publication date: August 2021.
Towards a consistent interpretation of AIOps models 13
Table 3. Description of metrics for Backblaze disk failure prediction.
Metric Description
Read Error Rate Frequency of errors while reading raw data from a disk.
Start/Stop Count 1Number of spindle start/stop cycles.
Reallocated Sectors Count Quantity of remapped sectors.
Seek Error Rate Frequency of errors while positioning.
Power-On Hours Number of hours elapsed in the power-on state.
Power Cycle Count Number of power-on events.
Reported Uncorrectable Errors
Number of reported uncorrectable errors, the denition is
vendor-specic.
Load Cycle Count Number of cycles into landing zone position.
HDA Temperature Temperature of a hard disk assembly.
Current Pending Sector Count Number of unstable sectors (waiting for remapping).
UltraDMC CRC Error Count Number of CRC errors during UDMA mode.
1
For cumulative SMART attributes, we extract both their raw value from the last day and the dierence during the
training period as features, following the setup of Lyu et al. [55].
And Regression Tree (CART), Gradient Boosting Decision Tree (GBDT), Random Forest (RF), and
Multi-layer Perceptron Neural Network (NN).
To mitigate the impact of dierent scales of metrics on model performance and interpretation,
we perform data standardization on each metric in the training dataset by removing the mean of
the metric and scaling the metric to its unit variance, using the
StandardScaler
function in the
scikit-learn Python package.
We observe that the dataset is extremely imbalanced, with only 1% failed jobs in the Google
dataset and 0.1% failed hard drives in the Backblaze dataset. To mitigate the impact of imbalanced
dataset on model performance, we downsample the majority class (i.e., succeed jobs in the Google
dataset and normal hard drives in the Backblaze dataset) in the training dataset to a success-to-fail
ratio of 10:1 prior to training the model.
The detailed conguration of the model training process is described in sections of each RQ.
4.2.3 Model evaluation. Prior work [
74
] shows that one should use threshold-independent metrics
such as Area Under the ROC Curve (AUC) in lieu of threshold-dependent metrics such as Precision,
Recall or F-measure to evaluate model performance. Therefore in this case study, we use AUC to
evaluate the performance of each trained model.
Because of the temporal nature of AIOps datasets, traditional evaluation methods such as cross
validation may result in data leakage and therefore lead to inaccurate performance evaluation [
55
].
In this case study, we evaluate a model that is trained on a specic period of dataset using data in
the next period as the testing dataset.
To prepare the testing dataset, we apply the same data scaler that was tted on the training
dataset on the testing dataset. It is worth noting that although we rebalance the training dataset by
downsampling the majority class, we do not perform such downsampling on testing dataset.
4.2.4 Model interpretation. There are two types of model interpretation: model-level interpretation
(i.e., identifying the features that have the biggest impact on a model’s predictions), and instance-
level interpretation (i.e., identifying the reason that a model predicts a specic input to be a specic
outcome). In this case study, we focus on model-level interpretation because prior work on AIOps
mostly focus on model-level interpretation [8, 16, 35, 40–42].
In particular, we use permutation feature importance, a model-agnostic approach to derive the
importance of features in a machine learning model in our study. Permutation feature importance
evaluates the importance of a feature by randomly shuing the value of the feature on the testing
ACM Trans. Softw. Eng. Methodol., Vol. , No. , Article . Publication date: August 2021.
14 Yingzhe Lyu, Gopi Krishnan Rajbahadur∗, Dayi Lin∗, Boyuan Chen∗, and Zhen Ming (Jack) Jiang
dataset, and measure the drop of model performance due to such shuing. We use permutation
feature importance in our case study because: (1) some of our studied learners are not intrinsically
interpretable; (2) permutation feature importance oers a consistent way to compare across all the
models; and (3) many existing studies use permutation feature importance on the interpretations
of machine learning models [
66
,
75
]. To control the impact of randomness in permutation feature
importance calculation on our case study results, we x the random seed of the permutation feature
importance calculation.
In the rest of the paper, we will use the term interpretation and feature importance ranks
interchangeably.
4.2.5 Similarity measurement of model interpretation. To compare the similarity of interpretations
among two or more models, we use the following measurements:
•Kendall’s Tau
: a non-parametric measure of the similarity between two rankings. Kendall’s
Tau ranges from 0 (no agreement) to 1 (complete agreement). We use the same interpretation
schema from prior work [66] to interpret Kendall’s Tau in our study:
Kendall’s Tau Agreement =
Weak if 0≤𝑇𝑎𝑢 ≤0.3.
Moderate if 0.3<𝑇𝑎𝑢 ≤0.6.
Strong if 0.6<𝑇𝑎𝑢 ≤1.
•Kendall’s W
: a non-parametric measure of the similarity among multiple rankings. Similar
to Kendall’s Tau, Kendall’s W also ranges from 0 (no agreement) to 1 (complete agreement).
We use the same interpretation schema that we use for Kendall’s Tau.
•Top K Overlap Score
: after ltering out features with a negligible importance score (i.e.,
permutation feature importance score lower than 0.0001), we use the same denition of the
Top K Overlap Score as prior work [66]:
Top K Overlap Score =
∩𝑛
𝑖≥2Most important K features for model 𝑖
∪𝑛
𝑖≥2Most important K features for model 𝑖
where 𝑛is the number of models for comparison.
5 (RQ1) ARE THE INTERPRETATIONS FROM AIOPS MODELS INTERNALLY
CONSISTENT?
In this section, we evaluate the internal consistency of AIOps model interpretation.
5.1 Approach.
Figure 1 shows three components involved in the model training phase of AIOps pipeline: data
sampling, hyperparameter tuning, and ML learner. Prior work [
43
,
64
] shows that randomness
in the model training phase leads to inconsistent model performance. In this RQ, we hypothesize
that randomness in the model training phase also leads to inconsistent model interpretations. In
particular, the tting algorithm of the ML learner may have inherent randomness; the randomized
hyperparameter tuning contains randomness; and the random sampling of training dataset also
introduces randomness to the model training phase [64].
To validate our hypothesis, we conduct a set of controlled experiments in this RQ: for each of
the potential sources of randomness, we control for the other sources of randomness and train the
7 studied learners on the two datasets. More specically:
(1)
To evaluate the impact of the inherent randomness from the learner on the internal consis-
tency of AIOps model interpretations, we control for the hyperparameters and the training
dataset, i.e., we x the hyperparameters of each learner to its default hyperparameters, and
ACM Trans. Softw. Eng. Methodol., Vol. , No. , Article . Publication date: August 2021.
Towards a consistent interpretation of AIOps models 15
Table 4. Controlled experiment setup for RQ1.
Experiment Potential source of randomness
ML learner Hyperparameter tuning Data sampling
1# ! !
2! # !
3! ! #
4! ! !
!- Controlled; #- Not Controlled.
train a model for each studied learner using all data in a given time period and dataset,
without controlling for the random seed of the learner’s tting algorithm. The process is
repeated 10 times, and the interpretation similarity measurements among the 10 iterations
are calculated for each learner on each dataset and its each time period.
(2)
To evaluate the impact of the randomness from randomized hyperparameter tuning on the
internal consistency of AIOps model interpretations, we control for the inherent randomness
from the learner and the training dataset, i.e., for each studied learner, we train a model
with randomized search on optimal hyperparameters, using all data in a given time period
and dataset, with a xed random seed for the learner’s tting algorithm. The process is
repeated 10 times for each dataset and its each time period, and the interpretation similarity
measurements among the 10 iterations are calculated for each learner on each dataset and its
each time period.
(3)
To evaluate the impact of the randomness from data sampling, we control for the inherent
randomness from the learner and the hyperparameters, i.e., for each studied learner, we x
the hyperparameters of the learner to its default hyperparameters, and the random seed for
the learner’s tting algorithm, and train a model using a bootstrapped sample in a given time
period and dataset. Unlike in Experiment 1, 3, and 4 where we use the same training data (i.e.,
all data in a given time period) for all iterations to control for training data randomness, here
through bootstrapping a dierent sample of training data in each iteration to intentionally
introduce data randomness. The process is repeated 10 times for each dataset and its each
time period, and the interpretation similarity measurements among the 10 iterations are
calculated for each learner on each dataset and its each time period.
(4)
Finally, to evaluate if we can derive an internally consistent interpretation of a AIOps model
when controlling for all the above three factors, we conduct a fully controlled experiment: we
x the hyperparameters of each learner to its default hyperparameters, and train a model for
each studied learner using all data in a given time period and dataset, with a xed random seed
for the learner’s tting algorithm. The process is repeated 10 times, and the interpretation
similarity measurements among the 10 iterations are calculated for each learner on each
dataset and its each time period.
Table 4 shows an overview of our controlled experiment setup.
5.2 Results.
Inherent randomness from learners introduces inconsistencies to the derived interpreta-
tions of an AIOps model.
Figure 2 shows the interpretation consistency among iterations when
controlling for the hyperparameters and the training dataset. As shown in Figure 2, learners such
as NN, RF and CART show inconsistent interpretation among iterations, while LDA, QDA, LR and
ACM Trans. Softw. Eng. Methodol., Vol. , No. , Article . Publication date: August 2021.
16 Yingzhe Lyu, Gopi Krishnan Rajbahadur∗, Dayi Lin∗, Boyuan Chen∗, and Zhen Ming (Jack) Jiang
BACKBLAZE
CART
BACKBLAZE
GBDT
BACKBLAZE
LDA
BACKBLAZE
LR
BACKBLAZE
NN
BACKBLAZE
QDA
BACKBLAZE
RF
GOOGLE
CART
GOOGLE
GBDT
GOOGLE
LDA
GOOGLE
LR
GOOGLE
NN
GOOGLE
QDA
GOOGLE
RF
0 10 20 30 0 10 20 30 0 10 20 30 0 10 20 30 0 10 20 30 0 10 20 30 0 10 20 30
0.25
0.50
0.75
1.00
0.25
0.50
0.75
1.00
Period
Value
measure
Kendall's W
Top 3 Overlap
Top 5 Overlap
Fig. 2. Consistency of the AIOps model interpretations across iterations under the impact of inherent
randomness from learners.
GBDT do not show inconsistent interpretation among iterations. It is worth noting that the degree
of inherent randomness of a learner’s tting algorithm depends on the specic implementation
of such algorithm. While scikit-learn’s implementation of LDA and QDA do not involve random
states
3
. Scikit-learn provides multiple solvers for LR, among which some involve randomness (
sag
,
saga
or
liblinear
) while others don’t
4
. Similarly, xgboost’s implementation of GBDT supports
multiple boosters and only shows non-deterministic behavior when a gblinear booster (i.e., Hogwild
algorithm) is used
5
. In our study, we used the default solver for LR (i.e.,
lbfgs
) and the default
booster for GBDT (i.e.,
gbtree
), both of which do not involve randomness. Hence, the apparent
stability in Figure 2 may be dependent on the implementation of the learners that we used and
may not be generalizable to other implementations.
Randomized hyperparameter searching introduces inconsistencies to the interpreta-
tion of AIOps models.
Figure 3 shows the interpretation consistency among iterations when
controlling for the inherent randomness from the learner and the training dataset. As shown in
Figure 3, randomized hyperparameter searching introduces inconsistent interpretation among
iterations across all studied learners.
Sampling randomness introduces inconsistencies to the interpretation of AIOps mod-
els.
Figure 4 shows the interpretation consistency among iterations when controlling for the
inherent randomness from the learner and the hyperparameters. As shown in Figure 4, sampling
(with bootstrap) introduces inconsistent interpretation among iterations across all studied learners.
When controlling all three sources of randomness, AIOps model interpretations are
internally consistent.
Figure 5 shows the interpretation consistency among iterations when
controlling for all three sources of randomness. As shown in Figure 5, all learners in every period
of both datasets yield identical interpretations across iterations. The result indicates that when
the randomnesses of learner, data sampling, and hyperparameter tuning are all controlled, AIOps
model interpretation becomes internally consistent.
The randomness from data sampling introduces the largest scale of inconsistency to
AIOps model interpretation.
Figure 6 provides an aggregated view of the distribution of Kendall’s
W when dierent sources of randomness are not controlled, in comparison to a controlled group
3https://github.com/scikit-learn/scikit-learn/blob/15a949460/sklearn/discriminant_analysis.py
4https://github.com/scikit-learn/scikit-learn/blob/15a949460/sklearn/linear_model/_logistic.py
5https://xgboost.readthedocs.io/en/latest/python/python_api.html
ACM Trans. Softw. Eng. Methodol., Vol. , No. , Article . Publication date: August 2021.
Towards a consistent interpretation of AIOps models 17
BACKBLAZE
CART
BACKBLAZE
GBDT
BACKBLAZE
LDA
BACKBLAZE
LR
BACKBLAZE
NN
BACKBLAZE
QDA
BACKBLAZE
RF
GOOGLE
CART
GOOGLE
GBDT
GOOGLE
LDA
GOOGLE
LR
GOOGLE
NN
GOOGLE
QDA
GOOGLE
RF
0 10 20 30 0 10 20 30 0 10 20 30 0 10 20 30 0 10 20 30 0 10 20 30 0 10 20 30
0.00
0.25
0.50
0.75
1.00
0.00
0.25
0.50
0.75
1.00
Period
Value
measure
Kendall's W
Top 3 Overlap
Top 5 Overlap
Fig. 3. Consistency of the AIOps model interpretations across iterations under the impact of randomized
hyperparameter searching.
BACKBLAZE
CART
BACKBLAZE
GBDT
BACKBLAZE
LDA
BACKBLAZE
LR
BACKBLAZE
NN
BACKBLAZE
QDA
BACKBLAZE
RF
GOOGLE
CART
GOOGLE
GBDT
GOOGLE
LDA
GOOGLE
LR
GOOGLE
NN
GOOGLE
QDA
GOOGLE
RF
0 10 20 30 0 10 20 30 0 10 20 30 0 10 20 30 0 10 20 30 0 10 20 30 0 10 20 30
0.00
0.25
0.50
0.75
1.00
0.00
0.25
0.50
0.75
1.00
Period
Value
measure
Kendall's W
Top 3 Overlap
Top 5 Overlap
Fig. 4. Consistency of the AIOps model interpretations across iterations under the impact of sampling
randomness.
BACKBLAZE
CART
BACKBLAZE
GBDT
BACKBLAZE
LDA
BACKBLAZE
LR
BACKBLAZE
NN
BACKBLAZE
QDA
BACKBLAZE
RF
GOOGLE
CART
GOOGLE
GBDT
GOOGLE
LDA
GOOGLE
LR
GOOGLE
NN
GOOGLE
QDA
GOOGLE
RF
0 10 20 30 0 10 20 30 0 10 20 30 0 10 20 30 0 10 20 30 0 10 20 30 0 10 20 30
0.950
0.975
1.000
1.025
1.050
0.950
0.975
1.000
1.025
1.050
Period
Value
measure
Kendall's W
Top 3 Overlap
Top 5 Overlap
Fig. 5. Consistency of the AIOps model interpretations across iterations when controlling all three sources of
randomness.
ACM Trans. Softw. Eng. Methodol., Vol. , No. , Article . Publication date: August 2021.
18 Yingzhe Lyu, Gopi Krishnan Rajbahadur∗, Dayi Lin∗, Boyuan Chen∗, and Zhen Ming (Jack) Jiang
BACKBLAZE
CART
BACKBLAZE
GBDT
BACKBLAZE
LDA
BACKBLAZE
LR
BACKBLAZE
NN
BACKBLAZE
QDA
BACKBLAZE
RF
GOOGLE
CART
GOOGLE
GBDT
GOOGLE
LDA
GOOGLE
LR
GOOGLE
NN
GOOGLE
QDA
GOOGLE
RF
0.4
0.6
0.8
1.0
0.4
0.6
0.8
1.0
Kendall's W
Source of Randomness (left to right): None ML learner Hyperparameter tuning Data sampling
Fig. 6. The interpretation consistency (measured by Kendall’s W) of models when dierent sources of
randomness are not controlled. Each subplot has four boxplots, from le to right represents the control
group (all sources of randomness controlled), randomness from ML learner, hyperparameter tuning, and data
sampling. See Table 4 for detailed setup.
with all sources of randomness controlled. It can be observed from Figure 6 that generally when
randomness from data sampling is introduced, the interpretation of AIOps models show the lowest
consistency; followed by hyperparameter tuning which shows the second lowest consistency
among model interpretations. The ML learner’s inherent randomness appears to introduce the
least inconsistency to AIOps model interpretation. Such observation is consistent across most
learners and datasets, except for Multi-layer Perceptron Neural Network (NN) - in the case of
NN, hyperparameter tuning introduces comparable interpretation inconsistency to data sampling.
A possible explanation is that for NN, the architecture of the network (e.g., the size of hidden
layers) are tuned as a group of hyperparameters, which may result in signicant dierence in the
architecture of the trained models. Because of the wide adoption of Neural-Network-based learners
in AIOps research (e.g., Deep Neural Network, DNN), future work is needed to further explore the
impact of hyperparameter tuning randomness on DNN model interpretation.
: Summary of RQ1
Inherent randomness from learners, randomized hyperparameter searching, and sampling
randomness all result in AIOps models yielding dierent interpretations across dierent
repeated runs. Sampling randomness introduces the largest scale of inconsistency to AIOps
model interpretation. By controlling the randomness from the learner, hyperparameter
tuning, and data sampling, we are able to derive internally consistent interpretations for
AIOps models.
6 (RQ2) ARE THE INTERPRETATIONS FROM AIOPS MODELS EXTERNALLY
CONSISTENT?
In this section, we evaluate the external consistency of AIOps model interpretation.
ACM Trans. Softw. Eng. Methodol., Vol. , No. , Article . Publication date: August 2021.
Towards a consistent interpretation of AIOps models 19
GOOGLE
k_2
k_3
k_4
k_5
k_6
k_7
k_8
k_9
k_10
k_11
k_12
k_13
k_14
k_15
k_2
k_3
k_4
k_5
k_6
k_7
k_8
k_9
k_10
k_11
k_12
k_13
k_14
k_15
0.0
0.1
0.2
0.3
0.4
Number of Clusters
WSS
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
BACKBLAZE Period
Fig. 7. Elbow graph of performance clustering for deciding the optimal number of clusters.
6.1 Approach.
Our approach to evaluate the external consistency of AIOps model interpretation consists of three
steps: (1) model generation; (2) model clustering; and (3) interpretation comparison.
Step 1: Model generation.
In Section 5, we nd that the randomness existing in hyperparameter
tuning, data sampling, and ML learners themselves would lead to a large variety of models being
generated. To preserve the randomness and generate models in dierent performance scales, we
train models on bootstrapped dataset samples, with randomized hyperparameter searching. For
each of the learners at each period, we repeat the model training process for 10 iterations. During
each iteration, a bootstrapped dataset sample is randomly generated without setting explicit random
seeds. We apply this process to both studied datasets. In the end, for each dataset and each period,
we generated 70 models. For example, the Google dataset contains data of 28 periods, hence we
generated 27
×
70
=
1
,
890 models. We did not use the data from the last period for training as there
would be no available data for testing the performance of models.
Step 2: Model clustering.
In this step, we cluster the models trained with the data in same
periods based on the model performance (i.e., the AUC metric). We use a widely-adopted one-
dimensional clustering technique called Jenks natural breaks optimization [
34
]. To decide the
optimal number of clusters, we rst conducted the elbow method. The elbow method [
78
] visualizes
the relationship between the number of clusters and the evaluation metric. It is a heuristic commonly
used in cluster analysis to decide the optimal number of clusters. The elbow method considers
the number of clusters to be optimal, if increasing the number of clusters has little impact on the
evaluation metrics. In this study, we use Within Sum of Square (WSS) as the evaluation metric.
WSS is to measure the variability of observations in each cluster and is calculated based on the
clustering technique (i.e., Jenks natural breaks optimization in our case). As the number of clusters
increases, the WSS will decrease as the variation in each cluster decreases. The optimal cluster
number is the cuto point where increasing cluster number has little impact on WSS.
Figure 7 shows the relations between WSS and the number of clusters for both Google and
Backblaze datasets. The x-axis shows the parameters we use in the Jenk natural breaks optimization.
For example,
𝑘_
2represents the nal number of cluster is one. From the visualization, we decide
that the optimal number of clusters for the Google dataset is 4 (i.e.,
𝑘_
5is the elbow point) and the
optimal number of clusters for the Backblaze dataset is 2 (i.e., 𝑘_3is the elbow point).
We then apply Jenks natural breaks optimization to the 70 models in each period of each dataset,
with the chosen optimal number of clusters, to group the models into dierent clusters based on
ACM Trans. Softw. Eng. Methodol., Vol. , No. , Article . Publication date: August 2021.
20 Yingzhe Lyu, Gopi Krishnan Rajbahadur∗, Dayi Lin∗, Boyuan Chen∗, and Zhen Ming (Jack) Jiang
their performance. We noticed that Jenks natural breaks cannot divide the models in Backblaze
period 12 into two clusters with at least 2 models in each cluster. We therefore removed Backblaze
period 12 from our analysis in the rest of this RQ.
Step 3: Interpretation comparison.
In this step, we calculate the interpretation similarity
measurements within each of the clusters to check if the interpretations of models in the same
performance group are externally consistent. As described in Section 4, we use three types of
similarity metrics to measure the consistency of interpretations: the Kendall’s W, Top 3 overlap
score, and Top 5 overlap score.
To statistically compare the interpretation consistency in dierent performance groups, we
perform the Wilcoxon Rank Sum (WRS) test. The Wilcoxon Rank Sum test is an unpaired, non-
parametric test commonly used in literature [
15
,
45
,
46
], of which the null hypothesis is that for
randomly selected values X and Y from two distributions, the probability of X being greater than Y
is equal to the probability of Y being greater than X. In this RQ, to conrm if the interpretation
consistency of group X is statistically signicantly higher than group Y, we use a one-sided
alternative hypothesis of X being shifted to the right of Y. Therefore, if the p-value of the Wilcoxon
Rank Sum test is less than 0.05, we conclude that group X has more consistent interpretations
within the group than group Y.
To mitigate the threat of false positives during multiple comparisons, we use Bonferroni correc-
tion [59] to correct the p-value before concluding a Wilcoxon Rank Sum test where applicable.
The Wilcoxon Rank Sum test only shows whether two distributions are dierent, but not the
magnitude of the dierence. To assess the magnitude of the dierences, we also calculate the eect
size using Cli’s Delta
𝑑
. We use the following schema for interpreting
𝑑
, which is widely used in
prior work [15, 45, 46]:
Cli’s Delta 𝑑=
Negligible (N) if 0≤ |𝑑| ≤ 0.147.
Small (S) if 0.147 <|𝑑| ≤ 0.33.
Medium (M) if 0.33 <|𝑑| ≤ 0.474.
Large (L) if 0.474 <|𝑑| ≤ 1.
6.2 Results.
RF and GBDT are the top two best-performing learners for the studied AIOps context.
Figure 8 shows the percentage of models trained with each learner in the best-performing clusters.
For the Google dataset, Figure 8 shows that RF and GBDT perform much better than other learners.
The median percentage of RF and GBDT are around 50% and the percentage of other ve learners
are all below 1%. Similarly, the results on the Backblaze dataset show that RF and GBDT are still
the top two learners among all seven learners. Dierent from the results on the Google dataset,
the performance of models trained using LDA, QDA, LR, and NN are comparable to those trained
using RF and GBDT. In general, none of the other ve learners generate more than 25% of models
in the best-performing cluster.
The ndings suggest that we should consider using RF and GBDT as the preferred learners in the
context of studied AIOps applications and datasets, as they consistently perform the best among all
studied learners. In addition, the performance of learners may vary due to dierent characteristics
of dierent datasets. Future work should further explore the correlation between the AIOps dataset
characteristics and performance of dierent learners.
Models in the highest-performed clusters tend to have more consistent interpretation
results compared to lower-performed clusters.
Figure 9 shows the three types of similarity
measurements for clusters in dierent performance ranks.
ACM Trans. Softw. Eng. Methodol., Vol. , No. , Article . Publication date: August 2021.
Towards a consistent interpretation of AIOps models 21
GOOGLE
RF GBDT QDA NN LR LDA CART RF GBDT LDA QDA LR NN CART
0
25
50
75
100
Learner
% in best-performing cluster
BACKBLAZE
Fig. 8. Percentage of models trained with each learner in the best-performing clusters.
For the Google dataset, the top 1 ranked cluster has higher values in all the three measurements
compared to other clusters. For example, the median of Kendall’s W in top 1 ranked cluster is
0.81, while the median of Kendall’s W in other clusters is only between 0.47 and 0.53. Table 5
presents the detailed results of the comparison. For all three types of measurements, the values of
the cluster in rank 1 are statistically higher than the clusters in rank 2, 3, and 4 (corrected p-value
<0.05). Furthermore, the results show that all the eect sizes are large. We also observe that the
highest-performed cluster in the Google dataset have much more consistent interpretations.
For the Backblaze dataset, there are only two performance clusters: rank 1 and rank 2. We observe
that, for the measurements of top 5 overlap score and top 3 overlap score, the dierences between
rank 1 cluster and rank 2 cluster are signicant dierent (corrected p-value <0.05). The magnitude
of dierences are either small or medium. On the other hand, for the measurement of Kendall’s
W, we nd that the dierence is insignicant. We further investigate the reason for insignicant
results of Kendall’s W. We notice that the AUCs of the models trained on Backblaze dataset are
mostly above 0.75. Based on such observations, we hypothesize that the insignicant dierence
may be due to the fact that all the trained models have relatively high performance. To verify our
hypothesis, we further explore the relationship between performance (i.e., AUC) and within-cluster
interpretation similarity measurements (e.g., Kendall’s W).
Models that have an AUC of at least 0.75 tend to have high consistency among their in-
terpretation.
Figure 10 shows the relationship between a cluster’s median AUC value and its
within-cluster interpretation consistency represented by Kendall’s W value. The blue line is tted
using a Local Polynomial Regression and the grey area represents a 95% condence interval. The
red line indicate the threshold of Kendall’s W (0.6) where the consistency is considered to be high.
Note that for the Backblaze dataset, all clusters have a median AUC above 0.70, with the majority of
them above 0.75, while for Google dataset, the median AUCs span across 0.5 to 0.95. For the gure
of the Google dataset in Figure 10a, we observe that the within-cluster interpretation similarity
begins to increase after AUC is over 0.7, and exceeds a Kendall’s W of 0.6 when AUC is larger than
0.75.
From Figure 10b, we observe that most of the clusters’ median AUC values are higher than 0.75
and Kendall’s W is consistently above the high agreement threshold. While it is unclear whether
the lower AUC values can still result in high Kendalls’W (i.e., > 0.6), from the observable trends in
both datasets, we draw a conservative conclusion that models that have an AUC of at least 0.75 tend
ACM Trans. Softw. Eng. Methodol., Vol. , No. , Article . Publication date: August 2021.
22 Yingzhe Lyu, Gopi Krishnan Rajbahadur∗, Dayi Lin∗, Boyuan Chen∗, and Zhen Ming (Jack) Jiang
GOOGLE
Kendall's W Top 3
Overlap
Score
Top 5
Overlap
Score
Kendall's W Top 3
Overlap
Score
Top 5
Overlap
Score
0.00
0.25
0.50
0.75
1.00
Measurement
Value
Performance
rank of the
cluster
1
2
3
4
BACKBLAZE
Fig. 9. Distribution of similarity measurements for clusters with dierent performance ranks.
Table 5. Comparing the similarity measurement among clusters in dierent performance ranks.
Measurement Dataset Rank X Rank Y Corrected p-value
for WRS+Cli’s 𝑑∗
Kendall’s W Google
1 2 < 0.01 0.71 (L)
1 3 < 0.01 0.84 (L)
1 4 < 0.01 0.89 (L)
Backblaze 1 2 0.91 - (N/A)
Top 5 Overlap Score Google
1 2 < 0.01 0.73 (L)
1 3 < 0.01 0.69 (L)
1 4 < 0.01 0.74 (L)
Backblaze 1 2 0.01 0.32 (S)
Top 3 Overlap Score Google
1 2 < 0.01 0.59 (L)
1 3 < 0.01 0.55 (L)
1 4 < 0.01 0.60 (L)
Backblaze 1 2 < 0.01 0.45 (M)
+𝑝<0.05 - Signicant; 𝑝≥0.05 - Insignicant.
∗N/A - Not Applicable; N - Negligible; S - Small; M - Medium; L- Large.
to have high consistency among their interpretations, and their interpretations are more reliable
than models with AUCs less than 0.75.
: Summary of RQ2
RF and GBDT are the top two best-performing learners for the studied AIOps context.
Models in the best performing cluster tend to have more externally consistent interpretations
compared to other groups. When the AUC scores are greater than 0.75, the interpretations
of AIOps models are externally consistent.
ACM Trans. Softw. Eng. Methodol., Vol. , No. , Article . Publication date: August 2021.
Towards a consistent interpretation of AIOps models 23
GOOGLE
0.5 0.6 0.7 0.8 0.9 1.0 0.5 0.6 0.7 0.8 0.9 1.0
0.00
0.25
0.50
0.75
1.00
Median AUC of the cluster
Kendall's W
BACKBLAZE
(a)
GOOGLE
0.5 0.6 0.7 0.8 0.9 1.0 0.5 0.6 0.7 0.8 0.9 1.0
0.00
0.25
0.50
0.75
1.00
Median AUC of the cluster
Kendall's W
BACKBLAZE
(b)
Fig. 10. Relationship between a cluster’s median AUC value and its within-cluster interpretation similarity,
measured by Kendall’s W. The red dashed line marks a Kendall’s W of 0.6 (a strong agreement), while the
green solid line marks an AUC of 0.75. The blue curve is fied using a Local Polynomial Regression and the
grey area represents a 95% confidence interval of the fit.
Algorithm 1 Streaming Ensemble Algorithm (SEA)
1: 𝑛𝑜_𝑜 𝑓 _𝑙𝑒𝑎𝑟 𝑛𝑒𝑟 𝑠 _𝑖𝑛_𝑒𝑛𝑠𝑒𝑚𝑏𝑙𝑒 ←− 0
2: 𝐸←− ∅ 𝐸is the ensemble
3: 𝑡𝑖𝑚𝑒 _𝑝𝑒𝑟𝑖𝑜𝑑𝑠 ←− 𝑡𝑖𝑚𝑒_𝑝𝑒𝑟𝑖𝑜𝑑𝑠 _𝑖𝑛_𝑎_𝑔𝑖𝑣𝑒𝑛_𝑑𝑎𝑡𝑎𝑠𝑒𝑡
4: 𝑛←− 𝑛𝑢𝑚𝑏𝑒𝑟_𝑜 𝑓 _𝑡𝑖𝑚𝑒 _𝑝𝑒𝑟𝑖𝑜𝑑𝑠
5: 𝑒𝑛𝑠𝑒𝑚𝑏𝑙 𝑒_𝑠𝑖𝑧𝑒 ←− 𝑛/2
6: while 𝑘in 𝑡𝑖𝑚𝑒 _𝑝𝑒𝑟𝑖𝑜𝑑𝑠 do
7: Train learner 𝐿ion 𝑘i
8: 𝑐𝑎𝑛𝑑𝑖𝑑𝑎𝑡𝑒 ←− 𝑀𝑒𝑎𝑛𝑆𝑞𝑢𝑎𝑟 𝑒𝑑𝐸 𝑟𝑟𝑜 𝑟 (𝐿i-1, 𝑘i)
9: if 𝑛𝑜_𝑜 𝑓 _𝑙𝑒𝑎𝑟 𝑛𝑒𝑟 𝑠 _𝑖𝑛_𝑒𝑛𝑠𝑒𝑚𝑏𝑙𝑒 ≤𝑒𝑛𝑠𝑒𝑚𝑏𝑙𝑒_𝑠𝑖𝑧𝑒 then
10: Add 𝐿i-1 to 𝐸
11: else
12: while 𝑒in 𝐸do
13: if 𝑐𝑎𝑛𝑑𝑖𝑑𝑎𝑡𝑒 <𝑀𝑒𝑎𝑛𝑆 𝑞𝑢𝑎𝑟𝑒𝑑 𝐸𝑟𝑟 𝑜𝑟 (𝑒j, 𝑘i)then
14: replace 𝑒jwith 𝑐𝑎𝑛𝑑𝑖𝑑𝑎𝑡𝑒
15: end if
16: end while
17: end if
18: end while
7 (RQ3) ARE THE INTERPRETATIONS FROM AIOPS MODELS CONSISTENT ACROSS
TIME (I.E., TIME CONSISTENT)?
In this section, we evaluate the time consistency of AIOps model interpretation.
7.1 Approach.
To understand how the temporal nature of AIOps data aects the generalizability of the derived
interpretation of models updated with dierent approaches, we rst divide the studied datasets into
multiple time periods. We divide the Google dataset into 28 one-day time periods and the Backblaze
dataset into 36 one-month time periods as we outline in Section 4.2.1. We then simulate the event of
ACM Trans. Softw. Eng. Methodol., Vol. , No. , Article . Publication date: August 2021.
24 Yingzhe Lyu, Gopi Krishnan Rajbahadur∗, Dayi Lin∗, Boyuan Chen∗, and Zhen Ming (Jack) Jiang
updating the constructed AIOps models to predict the outcome when the last time period becomes
available (i.e., the 28th and 36th time period of Google and Backblaze datasets respectively). Unlike
previous RQs, here we consider the available data from all the time periods (prior to the last time
period) to update the AIOps model. We consider four approaches for updating the AIOps models
including two periodic model retraining approaches (i.e., Sliding Window approach and Full History
approach) and two time-based ensemble approaches (i.e., Accuracy Weighted Ensemble (AWE) and
Streaming Ensemble Algorithm (SEA)). Finally, we compare if the derived interpretation of the
AIOps models updated with these studied approaches capture the historic trends present in the
data through the following three steps: 1) Test model generation, 2) Ground truth extraction, and 3)
Interpretation comparison.
1 2 3 ... n/2 n/2+1 n/2+2 ... n-1 n
Periods
SWEAWE FH SW
Choose the best performing model on the time period n
For each of the 7 studied learners 10 iterations
AWE FH
SWE SW
70 models 70 models 70 models 70 models
Test models
For each of the studied dataset
Derive interpretations from the test models
Training data for learners updated with SW
Testing data
AWE SWE FH SW
Model
interpretations
Fig. 11. Overview of Step 1: Test model generation. FH stands for Full History, SEA stands for Streaming
Ensemble Algorithm, AWE stands for Accuracy Weighted Ensemble, and SW stands for Sliding Window.
Step 1: Test model generation.
In this step, we rst construct the AIOps models (whose in-
terpretation we test) using data points from all the time periods (except the last one) to predict
the outcome of data points on the last time period of the studied datasets. Figure 11 presents
an overview of our test model generation step. We construct these AIOps models with all the 7
learners we outline in Section 4.2.2 and update them using the four model updating approaches. For
each learner, we update the associated models with one of the studied model updating approach,
the process is repeated for 10 iterations (with bootstrap sampling), resulting in 70 models being
constructed. We do so across both the studied datasets. Among the 70 generated models for each
model updating approach, we choose the best performing model (i.e., the model with the highest
AUC on the last time period of each of the studied dataset a.k.a test dataset). We then derive the
interpretation of the chosen best performing model (i.e., the test model). At the end, we obtain four
derived interpretations from four test models, each from one of the four studied model updating
ACM Trans. Softw. Eng. Methodol., Vol. , No. , Article . Publication date: August 2021.
Towards a consistent interpretation of AIOps models 25
approaches. We describe how each of the studied model updating approach leverages the historic
data and updates the model below.
Periodic model retraining approaches.
These approaches typically retrain the model whenever
data from a new time period becomes available. In our case study, we update the studied AIOps
models to predict on the 28th and 36th time period of Google and Backblaze datasets respectively.
•
Sliding window approach (SW): When updating the AIOps model with a SW approach, we
retrain the model with n/2 time periods (i.e., half of the available time periods). For instance,
when updating an AIOps model to predict the outcome on 28
th
time period in Google dataset,
we use the last 14 time periods to retrain the model.
•
Full history approach (FH): We use all the data available to retrain the AIOps model. For
instance, when retraining an AIOps model to predict the outcome on 28
th
time period in
Google dataset, we use all the last 27 time periods.
The time-based ensemble approaches.
Instead of updating a single AIOps model, the time-based
ensembles combines several AIOps models trained on short periods of time (e.g., models trained
on each time period (local models)) in an ensemble [
73
,
80
]. We then use this ensemble model to
predict the outcome on future instances. For both the ensemble approaches that we use in this
study, we set the ensemble size (i.e., the number of local models that we ensemble together) at n/2
similar to the SW approach. Below we describe the two ensemble approaches that we consider in
our study.
•
Streaming Ensemble Algorithm (SEA) combines multiple local models using a majority vote.
Algorithm 1 presents the working of the SEA. To update the ensemble, SEA replaces the
weakest model in the ensemble by observing which of the local models perform the worst in
the latest time period for which the SEA is updated. The SEA ensembles multiple learners as
follows: When data points from a time period kbecomes available, a new learner L
i
(e.g., RF
learner) is trained on these data points to build a local model (please see line 7 in Algorithm 1).
Then, these data points are used to evaluate the model L
i-1
trained on the previous time
period k
i-1
(please see line 8 in Algorithm 1). If the number of models in the ensemble hasn’t
reached the predetermined ensemble size (n/2 in our case), the model L
i-1
is simply appended
to the ensemble (please see lines 9-10 in Algorithm 1). Else, the Mean Squared Error (MSE) of
the model L
i-1
is compared to all the models in the ensemble. If the MSE of the model L
i-1
is
lower than that of all the models in the ensemble, the model L
i-1
replaces the weak learner, if
not, the model L
i-1
is discarded (please see lines 10-16 in Algorithm 1). When the next new
set of data points become available for training, the model L
i
pertaining to the current time
period becomes the one that is tested against the other models in the ensemble.
•
Accuracy weighted ensemble (AWE) is similar to the SEA except on two key points. First,
instead of using a majority vote, AWE assigned weights to each model in the ensemble (we
detail the weight assignment below). Second, when data from a new time period kbecomes
available for training, SEA determines if model trained in the previous time period (L
i-1
)
should be kept or removed from the ensemble. Whereas, in AWE, we evaluate if the model
(Li) should be kept or discarded.
The weight for each model in the ensemble is assigned by calculating its prediction error on the
latest training data time period k
i
. More specically, we calculate the Mean Square Error (MSE) of
each model in the ensemble on the latest training data time period k
i
. The MSE for each learner is
calculated by computing the dierence between the predicted probability of the given model and
the observed outcome class (either 0 or 1) in k
i
. We note it as MSE
i
. The weight for each model
in the ensemble is then given by w
i
=MSE
r
-MSE
i
, where MSE
r
is the MSE value of a model that
ACM Trans. Softw. Eng. Methodol., Vol. , No. , Article . Publication date: August 2021.
26 Yingzhe Lyu, Gopi Krishnan Rajbahadur∗, Dayi Lin∗, Boyuan Chen∗, and Zhen Ming (Jack) Jiang
predicts randomly (i.e.,MSE
r
=0.25). The models in the ensemble that perform worse than the MSE
r
are removed from the ensemble. We use a 10-fold cross validation to estimate the MSE.
1 2 3 ... n/2-1 n/2 n/2+1 ... n-1 n
Periods
Candidates
Local AWE
SW SWE
FH
...
..
.
..
.
..
.
..
.
...
..
.
..
.
..
.
..
.
...
..
.
..
.
..
.
..
.
...
..
.
..
.
..
.
..
.
... ...
Ground truth candidates for
each period
Derive interpretation of the
best performing model in
each time period
Ground truth for each period
For each of the studied dataset
Ground truth not computed for time periods that are marked in red
350 models
Select the best performing model
in each time period
Gn/2
Derive interpretation of the best performing
model in each time period
Gn/2+1 Gn/2+x Gn-1
Fig. 12. Overview of Step 2: Ground truth extraction. FH stands for Full History, SEA stands for Streaming
Ensemble Algorithm, AWE stands for Accuracy Weighted Ensemble, and SW stands for Sliding Window.
Step 2: Ground truth extraction.
In this step, we extract the actual trends (i.e., ground truth)
encapsulated in each time period of the studied datasets. Figure 12 presents an overview of our
ground truth extraction step. We do so to verify if the derived interpretation of the test models
obtained from the previous step are faithful to the historic data. While it is impossible to accurately
know what features were the key drivers in prior time periods, as RQ2 results suggest, the inter-
pretation of models in the high performing clusters are generally consistent. Furthermore, several
prior studies also suggest that high performing models can typically be interpreted reliably [
49
,
66
].
Therefore, to approximate the ground truth of each time period, we pick the best performing model
in that time period and obtain its derived interpretation. For instance, to extract the ground truth
of a given time period n-1, we pick the model that is trained on n-1
th
time period (or on all the
time periods before it) and has the best performance on the n
th
time period. We then derive the
interpretation of this best performing model as a proxy for the ground truth contained in the
time period n-1. For example, in Google dataset, if we were to extract the ground truth for time
period 25, then we consider the local models trained on 25
th
time period and models trained on 25
th
time period that are updated with the studied AIOps model updating approaches. Among these
constructed models if we were to nd that the RF model updated with FH approach to be the best
performing model on the 26
th
time period, we then derive the interpretation of this model. We
consider this derived interpretation as the ground truth of the 25th time period.
ACM Trans. Softw. Eng. Methodol., Vol. , No. , Article . Publication date: August 2021.
Towards a consistent interpretation of AIOps models 27
We compute the ground truth from n/2 to n-1 (where n=#time periods) time periods for each
of the studied datasets by extracting the derived interpretation of the best performing model in
the given time period. We do not compute the ground truth on the n
th
time period as we are only
trying to observe how faithful the best performing model on n
th
time period is to the ground truth
available in time periods n/2 to n-1. Furthermore, instead of computing the ground truth on all the
time periods starting from 1, we do only from (n/2)
th
time period, as we include both the AIOps
models trained only on one time period (similar to RQ2) and AIOps models trained using the studied
model updating approaches. However, the SW approach requires 1 to n/2 time periods to train. For
instance, in Google dataset, for a learner to be updated with SW approach, the rst 13 time periods
are required to update the model trained on 14
th
time period. We include both the models trained
only on one time period and models trained using the studied model updating approaches since we
want to maximize the chance of nding the best performing model in each time period.
We train all the studied learners that we outline in Section 4.2.2 on time periods n/2 to n-1 of the
studied datasets. We rst train them using the same approach that we outline in RQ2. On each of
the time period between n/2 and n-1 we rst end up with 70 models. Next, we train our studied
learners with the four model updating approaches rst using time periods 1 to n/2 as the training
data to predict on the outcome on (n/2+1)
th
time period. We incrementally add each time period
from (n/2+1)
th
until n-1
th
time period to the training data. For instance, in the Google dataset, we
rst build AIOps models on the 14
th
time period, then we incrementally train models from the 14
th
to the 27
th
time period. Therefore, on each time period we end up with 280 models. Finally, among
these 280
+
70
=
350 ground truth candidate models constructed for each time period between n/2
to n-1 we compute the best performing model and interpret it. We consider derived interpretation
of the best performing model in each time period as ground truth for the given time period. We do
so because it best captures the variation of the dataset during that period.
AWE SWE FH SW
Gn/2 Gn/2+1 Gn/2+x Gn-1
Compute similarity in terms
of Kendall’s Tau
For each of the studied dataset
...
Fig. 13. Overview of Step 3: Test model generation. FH stands for Full History, SEA stands for Streaming
Ensemble Algorithm, AWE stands for Accuracy Weighted Ensemble, and SW stands for Sliding Window.
Step 3: Interpretation comparison.
To observe if the derived interpretations of the test models
are impacted by the temporal nature of the AIOps data we compare its derived interpretation
with the ground truth extracted in each prior time period. Figure 13 presents an overview of
our interpretation comparison step. We do so by computing the similarity between the derived
interpretation of test models (please note that four test models given by each studied model updating
approach) and ground truth of each historic time period in terms of Kendall’s Tau. We assert that
temporal nature of the AIOps data does not impact the derived interpretation of test models, if the
similarity scores given by Kendall’s Tau is consistently strong across all historic time periods (i.e.,
n/2th to n-1th time periods).
ACM Trans. Softw. Eng. Methodol., Vol. , No. , Article . Publication date: August 2021.
28 Yingzhe Lyu, Gopi Krishnan Rajbahadur∗, Dayi Lin∗, Boyuan Chen∗, and Zhen Ming (Jack) Jiang
GOOGLE
16
20
24
28
20
25
30
35
0.25
0.50
0.75
1.00
Period
Kendall's Tau
AWE
Full History
SEA
Sliding Window
BACKBLAZE
Model update
approach
Fig. 14. Similarity measurement (Kendall’s Tau) between the interpretation of studied AIOps modelling
approaches and estimated ground truth. The red dashed line marks a Kendall’s Tau of 0.6 (a strong agreement).
Next, we seek to observe if the derived interpretation of a test model exhibits higher similarity
to the ground truth across all the historic time periods compared to the others. If it does, the model
updating approach of the given test model can be thought of as being the most faithful to the
trends present in the prior data. To compute that, we conduct a paired Kruskall-Wallis H-test [
39
]
between the computed similarity scores of the test models across all the time periods. A p-value
≤
0
.
05 on the Kruskall-Wallis H-test indicates that at least one of test model produces interpretation
which has consistently higher similarity with the ground truth across the historic time periods. For
instance, best performing test model updated by SEA to predict the outcome on 28
th
time period
would have 14 similarity scores in the Google dataset. Similarly the best performing test model
updated by the other studied model updating approaches would also 14 similarity scores associated
with each of them. We would then compute a Kruskall-Wallis H-test between these four similarity
score distributions to see if the distribution associated with any one of the studied approaches is
higher than the others.
If the Kurskall-Wallis test indicates that at least one of the test model has a dierent similarity
score distribution than others on a given dataset, we also conduct a pairwise Wilcoxon-rank sum
test and Cli’s delta eect size test similar to RQ2. We do so in order to identify the best model
updating approach(es) that yield(s) a signicantly superior similarity with the ground truth in
comparison to the other studied model updating approaches.
7.2 Results.
The temporal nature of the AIOps data impact the similarity between the ground truth
and the derived interpretation of the AIOps models updated using the studied model
updating approaches.
From Figure 14, we observe that similarity between the ground truth and
derived interpretation of the test models uctuates across all the time periods for both studied
datasets. In particular, except for 3 instances, none of the AIOps models constructed with the
studied approaches have a similarity score of 1 (i.e., a perfect reection of the ground truth) with
the ground truth observed in the studied periods across both studied datasets.
The derived interpretation of AIOps models updated with SW approach and FH ap-
proach have a consistently higher similarity scores with the ground truth across all his-
toric time periods than models updated with other studied approaches.
Figure 14 shows
that across the studied time periods, derived interpretation of test models trained with SW and FH
ACM Trans. Softw. Eng. Methodol., Vol. , No. , Article . Publication date: August 2021.
Towards a consistent interpretation of AIOps models 29
Table 6. Comparing the dierence in similarity measurements (i.e., the similarity between the ground truth
interpretation and derived interpretation of a AIOps model updated with the given update approach) between
the derived interpretation for AIOps models updated with SW approach and the AIOps models updated with
the other studied approaches.
Update Approach Dataset
FH SEA AWE
Corrected
p-value+Cli’s 𝑑∗Corrected
p-value+Cli’s 𝑑∗Corrected
p-value+Cli’s 𝑑∗
SW Google 0.86 - (N/A) < 0.01 0.47 (L) < 0.01 0.69 (L)
BackBlaze 0.35 - (N/A) < 0.01 0.49 (L) 0.02 0.64 (L)
+𝑝<0.05 - Signicant; 𝑝≥0.05 - Insignicant.
∗N/A - Not Applicable; N - Negligible; S - Small; M - Medium; L- Large.
Table 7. Comparing the dierence in similarity measurements (i.e., the similarity between the ground truth
interpretation and derived interpretation of a AIOps model updated with the given update approach) between
the derived interpretation for AIOps models updated with FH approach and the AIOps models updated with
the other studied approaches.
Update Approach Dataset
SW SEA AWE
Corrected
p-value+Cli’s 𝑑∗Corrected
p-value+Cli’s 𝑑∗Corrected
p-value+Cli’s 𝑑∗
FH Google 2.19 - (N/A) 0.001 0.47 (L) 0.001 0.63 (L)
BackBlaze 2.68 - (N/A) 0.006 0.57 (L) 0.03 0.40 (L)
+𝑝<0.05 - Signicant; 𝑝≥0.05 - Insignicant.
∗N/A - Not Applicable; N - Negligible; S - Small; M - Medium; L- Large.
approaches have a higher similarity score with the ground truth than the two time-based ensembles.
In addition, we also observe that on Backblaze dataset, test models updated with AWE and SEA
approaches have a strong similarity with ground truth only on 4 out 18 and 1 out of 18 time periods
respectively.
The Kruskall-Wallis test on the computed agreements between the interpretations of the test
models and the ground truth resulted in signicant (p<0.05) across both datasets. Such a result
indicates that at least the similarity scores between the derived interpretation of the test models
updated with one the studied model updating approach is signicantly dierent from the similarity
measurements produced by the other studied approaches.
Table 6 and Tabel 7 presents the p-values of a pairwise Wilcoxon-rank sum test (p-value corrected)
between the similarity scores of test models updated with SW and FH approach against the computed
similarity scores of test models updated with the other studied approaches. From Table 6 and Tabel 7,
we nd that the derived interpretation of best performing test models trained with both the SW
model and the FH model produces signicantly higher similarity with the ground truth than
the test models updated with time-based ensembles approaches across both the studied datasets.
Furthermore, we nd that the magnitude of eect sizes is consistently large. However, we do not
observe signicant dierences between the similarity measurements of the SW and FH models
across either of the studied datasets.
ACM Trans. Softw. Eng. Methodol., Vol. , No. , Article . Publication date: August 2021.
30 Yingzhe Lyu, Gopi Krishnan Rajbahadur∗, Dayi Lin∗, Boyuan Chen∗, and Zhen Ming (Jack) Jiang
: Summary of RQ3
The temporal nature of the AIOps data indeed impacts the similarity between the interpre-
tation of the AIOps models and the ground truth. Among the commonly used approaches to
update AIOps models, we nd that the derived interpretations from AIOps models updated
with SW and FH approaches are the closest to the ground truth across all time periods.
8 DISCUSSIONS AND FUTURE WORK
In this section, we rst discuss potential alternative experiment setups that could be applicable to
our study, and then highlight the gap between AIOps studies and the reproducibility guidelines in
the machine learning community. We would like to emphasize that the goal of our study is not to
propose an optimal and improved AIOps model, but rather to reveal the threats or pitfalls of prior
work’s practices when interpreting AIOps models. Therefore we keep the setup of the study close
to prior work. Nevertheless, the alternative experiment setups discussed below can be examined in
future work to conrm the generalizability of our ndings under dierent setups.
8.1 Training setups
In this study, we make several experimental choices in the training setups that we use to train our
AIOps models. More specically,
(1) We choose to standardize the independent metrics before training a model;
(2) We downsample the training dataset prior to training the model;
(3) We do not use any synthetic class rebalancing method like SMOTE.
We do so for two reasons. First, as we mentioned earlier, we wish to keep our AIOps model
training setup as close to prior work as possible. All of the aforementioned experimental choices
were used in prior work in the eld of AIOps. For instance, Lyu et al. [
54
] standardized the
independent metrics of their input data before building their AIOps models. Similarly, several prior
studies in AIOps [11, 56] tend to use a downsampling strategy similar to ours.
Second, the experimental choices that we make to build our AIOps models are quite robust.
Though we tried to keep the setup of our study as close to prior work as possible, we took care to
ensure that the experimental choices that we make are not sub-optimal. As Thomas et al
. [77]
and
Bring
[13]
state, the derived interpretations of a model are not impacted by the standardization
methods such as StandardScaler. Similarly, we do not use advanced class rebalancing methods
SMOTE to rebalance our datasets. Several prior studies [
74
,
75
,
79
] show that rebalancing the
datasets with techniques like SMOTE shifts the distribution of training data and impacts the derived
interpretations. Since our study focuses on evaluating the consistency of interpretations derived
from AIOps models, we avoided experimental choices that may impact the derived interpretations of
a model. However, we highly encourage future research to study the impact of dierent experimental
choices on the consistency of the interpretation derived from AIOps models using the rigorous set
of criteria that we outline in our study.
8.2 Evaluation setups
In this study, we used AUC as the metric to evaluate the performance of models. As stated in
Section 4.2.3, we selected AUC as the evaluation metric because prior work [
74
] has shown that
threshold-independent metrics such as AUC provides more stable evaluation results than threshold-
dependent metrics such as precision, recall and F-measure.
ACM Trans. Softw. Eng. Methodol., Vol. , No. , Article . Publication date: August 2021.
Towards a consistent interpretation of AIOps models 31
In particular, we used Area Under ROC Curve (AUROC) in our study. ROC curve, or Receiver
Operating Characteristic curve, is a curve that describes the relationship between a model’s true
positive rate and false positive rate [
10
]. The AUROC value has a clear interpretation schema. An
AUROC value of 0.5 indicates a model performance that is equivalent to random guessing, while
an AUROC value of 1 indicates a model that can provide perfect prediction with 0 error rate under
a certain threshold. However, when a given dataset is extremely imbalanced (like in many AIOps
datasets), AUROC can potentially produce over-optimistic evaluation results.
An alternative metric that is insensitive to imbalanced data is Area Under Precision-Recall Curve
(AUPRC), which describes the relationship between a model’s precision and recall. However unlike
AUROC which has a xed baseline of 0.5, AUPRC does not have a xed baseline for comparison. In
contrast, the baseline to evaluate whether an AUPRC value is good depends on the ratio of positive
samples in the dataset [
12
]. In RQ2, we studied the relationships between model performance and
external consistency of the interpretation. To study such relationships, we compared and clustered
many dierent models that are not necessarily trained on the same sample of the dataset. In cases
such as those, it is not practical to use AUPRC in our study setup to conduct a fair comparison
across models (since AUPRC does not have a xed baseline for dierent datasets like AUROC). In
addition, using AUPRC would not let us suggest a general acceptable threshold as a guideline for
AIOps model interpretation.
8.3 Interpretation levels
In this study, we used a model-level, model-agnostic approach (i.e., the permutation feature impor-
tance) to interpret AIOps models, as prior work on AIOps mostly focus on model-level interpretation.
However, in some cases, an instance-level interpretation could be useful for improving the model
quality and enhancing the trustworthiness of AIOps models. For example, it is possible that, al-
though two instances have similar features, the prediction results dier (e.g., one predicted as a
successful job run and the other predicted as a failed job run). In such cases, it is important to
understand the rationale behind the predictions, which could provide insights on how to improve
the performance of the AIOps model in the future.
Recently, many instance-level interpretation techniques (e.g., LIME [
69
], Anchor [
70
], and
SHAP [
52
]) have been proposed in the research literature [
27
,
52
,
63
,
69
,
70
]. The general idea of
instance-level interpretation techniques is to rst use local surrogate models (e.g., a linear regression
model) to approximate the predictions made by complex models (e.g., a deep neural network model),
and then derive interpretations from the local surrogate models which are inherently interpretable.
There are two common types of derived interpretations for a prediction: the feature importance
ranks and a set of rules and depending on the need both of these can be useful. For instance, Zhao
et al. [
88
] applied LIME to generate a visualized report for AIOps engineers to understand each
incident prediction result.
While it is possible (and useful) to apply instance-level interpretation in AIOps models, a recent
study [
23
] in the malware detection domain evaluated ve instance-level interpretation techniques.
They found that the derived interpretations are not consistent even if the prediction is made by the
same model with the same input. Within the ve studied instance-level interpretation techniques,
LIME [
69
] achieves the most stable results when the model is slightly changed. Such ndings indicate
a need for a study similar to ours to provide guidelines to ensure consistent interpretations can be
obtained from AIOps models. However, since our study only focused on model-level interpretation,
it remains unknown whether our ndings apply to instance-level interpretations in general and we
suggest that future work should explore this in detail.
ACM Trans. Softw. Eng. Methodol., Vol. , No. , Article . Publication date: August 2021.
32 Yingzhe Lyu, Gopi Krishnan Rajbahadur∗, Dayi Lin∗, Boyuan Chen∗, and Zhen Ming (Jack) Jiang
8.4 The gap between reproducibility guidelines in the machine learning community
and AIOps studies
There is a well-accepted reproducibility checklist [
65
] proposed in the machine learning commu-
nity which outlines the requirements when submitting research manuscripts to support better
reproducibility of the important ndings. The checklist contains 21 guidelines which are listed
in the appendix of this paper. There are ve categories of the 21 guidelines: models/algorithms,
theoretical claims, datasets, shared code, and experimental results. In order to understand whether
the recent AIOps studies follow the guidelines proposed in the machine learning community, we
revisited the 11 previously surveyed AIOps studies in Section 3.2.2.
All the studies that we included in our survey have followed the models/algorithms related
guidelines presented in the reproducibility guideline. None of the surveyed studies propose any
theoretical claims, hence guideline No. 4 and No. 5 do not apply. For the guidelines related to
datasets, only 2 out of the 11 studies provided a downloadable link for the dataset that they use
in their study. 6 out of the 11 studies did not explain the detailed process on how they collect the
data if newly presented, against guideline No.10. For the guidelines related to shared code, 10 out
of 11 studies fail to provide training code, evaluation code, or (pre-) trained models, which goes
against guideline No. 12, 13, and 14. None of the surveyed studies provide precise commands in the
README le to reproduce the results, which goes against guideline No. 15. For the experimental
result part, 5 out of 11 studies did not mention the approaches that they use for selecting the
hyperparameters. 7 out of 11 studies did not specify the exact number of training and evaluation
runs.
Our results show that there is currently a gap between recent AIOps studies and the repro-
ducibility guidelines in the machine learning community. Even though it is understandable that the
guidelines are not explicitly followed (for example, for guideline No. 9, it is reasonable that certain
datasets cannot be disclosed due to company policies, and the guidelines were only introduced in
2019), it is concerning to nd that so many studies violate the guidelines for reproducibility. In
addition, even by following the reproducibility guidelines, the results might not be reproducible
due to uncontrolled randomness (as we show in RQ1). Hence, our rst guideline regarding the
internal consistency complements the reproducibility checklist. It is crucial for the future AIOps
studies to start accounting for both the reproducibility guidelines and our guidelines to address
reproducibility concerns and subsequently achieve the internal consistency of interpretations.
9 GUIDELINES TO PRACTITIONERS
In this section, based on our ndings, we recommend the following practical guidelines for AIOps
researchers and practitioners to reliably interpret their AIOps models
[Guideline 1]: Always nd ways to expose and record the random seeds used to control
the non-determinism from the learner, hyperparameter tuning, and data sampling when
building an AIOps model.
From the results presented in Section 5, we observe that identical and
internally consistent interpretations could be derived from the constructed AIOps models only
when common sources of non-determinism were controlled. It is important to ensure the internal
consistency of the derived interpretations of an AIOps model to enable managers and DevOps
engineers to act based on these interpretations. For instance, if a DevOps engineer gets a dierent
feature as the source of a problem every time the an AIOps model is retrained (even on the same
setup), then it would be impossible to act upon. More importantly, as Krishna and Menzies
[38]
outline, the DevOps engineer would lose trust in the constructed AIOps model when obtained
insights are not consistent.
ACM Trans. Softw. Eng. Methodol., Vol. , No. , Article . Publication date: August 2021.
Towards a consistent interpretation of AIOps models 33
[Guideline 2]: Only use AIOps models with a minimum acceptable performance (i.e., AUC
greater than 0.75) to derive interpretations.
From the results presented in Section 6, we note
that derived interpretations of models with low performance (i.e., AUC less than 0.75) exhibit very
low external consistency. Only when the model’s performance in terms of AUC is greater than 0.75
the derived interpretations among similarly performing models start to become consistent. In light
of these ndings, we caution against the usage of lower performing interpretable models in lieu of
higher performing ones for deriving interpretations [
42
]. Particularly since, interpretations derived
from such models may not accurately reect the ground truth.
[Guideline 3]: When using AIOps models to derive interpretations, to update them when
new data becomes available, use periodic retraining strategies like Sliding Window and
Full History approaches.
As we mention in Section 9, it is common for AIOps researchers and
practitioners to update their trained AIOps models to keep up with the constant evolution of the
AIOps data. When doing so, as we observe from the results presented in Section 7, it is possible
for the updated models to lose sight of the historic trends. Only the models updated with Sliding
Window and Full History approaches consistently exhibit high similarity with the historic ground
truth present in the data.
10 THREATS TO VALIDITY
In this section, we discuss the threats to validity of our case study.
10.1 Internal Validity
In this paper, we studied four dierent AIOps model update approaches and compared their
interpretation. The window size of the sliding window approach is set to be a half of the number of
total time period, which is consistent with prior work [55].
10.2 External Validity
In this paper, we used seven learners on two datasets and leveraged the permutation feature scores
for interpretation. The two datasets (the Google cluster trace dataset and the Backblaze hard drive
statistics dataset) and the seven learners are widely adopted in previous AIOps studies [
42
,
55
].
Although our results might not generalize when using other learners, datasets, and interpretation
methods, our case study setup is generic and can be applied to other types of predictive AIOps
tasks in future work.
For AIOps tasks that leverage other types of machine learning algorithms such as unsupervised
learning, active learning, or reinforcement learning, the generalizability of our ndings should be
examined in future work. In addition, our study only focuses on AIOps tasks. Hence, generalizability
of our guidelines should be examined in other domains.
In RQ2, we derive a conclusion that models with a minimum AUC value of 0.75 tend to have high
consistency among their interpretations. This conclusion is based on the empirical observations
made on the two studied datasets. For other datasets and tasks, it is possible that models that do
not reach the minimum AUC value threshold (i.e., 0.75) could also have high consistency among
their interpretations. Future work is needed to study the generalizability of our conclusions
We implemented our experiments using Python and the 0.24.0 version of scikit-learn API. We did
not investigate the impact of dierent implementations in other programming languages (e.g., R,
C++) and other popular machine learning frameworks such as Weka and tensorow. In addition, we
only used the permutation feature importance scores provided by the scikit-learn API to derive the
interpretations from each model while there are other implementations [
1
,
2
]. Although we think
ACM Trans. Softw. Eng. Methodol., Vol. , No. , Article . Publication date: August 2021.
34 Yingzhe Lyu, Gopi Krishnan Rajbahadur∗, Dayi Lin∗, Boyuan Chen∗, and Zhen Ming (Jack) Jiang
this technique ts best in our context, there might be other techniques available to quantitatively
derive the interpretations. Future study is encouraged to investigate this matter.
10.3 Construct Validity
In the setting of all RQs, we downsample the majority class in the training set so that the success-
to-fail ratio is 10:1, which may present a threat to our construct validity. However, this process is
consistent with prior work to mitigate the imbalance of the dataset [11, 56].
In the setting of RQ1, we repeat the model training process for at least 10 times for each learner
to observe the impact of randomness. Similarly, in RQ2, the performance threshold of producing
consistent interpretation might be impacted by the characteristics of the randomly sampled datasets.
We repeated the training task for each learner in each period of data for at least 10 times to increase
the variety of sampled dataset and trained models, and mitigate the risk.
To measure the consistency of the interpretations, we used three types of similarity measurements:
the Kendall’s W, top 5 overlap score, and top 3 overlap score. These measurements have been widely
used in previous studies[55, 66].
We applied an automated process (i.e., randomized search) for hyperparameter tuning in the
training process and set the iteration times to be 100. We choose this method over manually
tuning as it would be impractical to manually tune every model with randomly sampled data. In
addition, Bergstra and Bengio [
9
] show that randomized search is ecient in locating the optimal
hyperparameters. Even though we cannot guarantee that we extract the best set of hyperparameters,
the impact on our study results is limited, since our experiments are conducted on two datasets
and seven learners, and our ndings do not rely on the hyperparameters being the most optimal.
11 CONCLUSIONS
In this paper, we study the consistency of interpretation of AIOps models through a case study
on two popular AIOps use cases: (1) job failure prediction using the Google cluster trace dataset;
and (2) hard drive failure prediction using the Backblaze hard drive statistic dataset. We assess the
consistency of AIOps model interpretation along three dimensions: internal consistency, external
consistency, and time consistency. Our results show that inherent randomness from learners,
randomized hyperparameter searching, and sampling randomness can lead to internally inconsistent
interpretations of AIOps models. In addition, we observe that the performance of AIOps models
impact the external consistency of interpretations—the interpretations of AIOps models in the
highest-performing clusters tend to be more consistent than those of other models. Finally, we
nd that AIOps models built using Sliding Window and Full History approaches have the most
consistent interpretations to the evolving ground truths.
In light of these ndings, we suggest the AIOps researchers and practitioners to: 1) Always control
the non-determinism from the learner, hyperparameter tuning, and data sampling by exposing
and recording random seeds; 2) Use models with high performance (i.e., AUC greater than 0.75)
to derive interpretations; and 3) Use either a Sliding Window or Full History approach to update
the model when using the updated model to derive interpretations. Our ndings and guidelines
can help AIOps researchers and practitioners derive consistent interpretations from AIOps models.
For the AIOps applications such as issue prediction and the issue mitigation, by following our
guidelines, practitioners can derive more reliable and consistent interpretations that explain the
occurrence of failures or outages in the eld, which can help them make better managerial decisions
that have long-lasting eects and develop better tooling support that saves maintenance eorts.
In the future, we plan to extend our study to more types of AIOps tasks such as performance
anomaly detection and diagnosis and other types of interpretation techniques, such as instance-
level interpretations. In addition, future work can compare the interpretation between manually
ACM Trans. Softw. Eng. Methodol., Vol. , No. , Article . Publication date: August 2021.
Towards a consistent interpretation of AIOps models 35
tuned models and automated tuned models. We are also interested in applying quality assurance
techniques such as metamorphic testing and dierential testing to further verify the quality of the
model interpretations within the AIOps context.
REFERENCES
[1] 2021. ELI5. https://eli5.readthedocs.io/en/latest/overview.html [Website accessed on 02-19-2021].
[2]
2021. PIMP - R implementation for permutation importance score. https://www.rdocumentation.org/packages/vita/
versions/1.0.0/topics/PIMP [Website accessed on 02-19-2021].
[3]
Julius Adebayo, Justin Gilmer, Michael Muelly, Ian J. Goodfellow, Moritz Hardt, and Been Kim. 2018. Sanity Checks
for Saliency Maps. In Advances in Neural Information Processing Systems 31: Annual Conference on Neural Information
Processing Systems 2018, NeurIPS 2018, December 3-8, 2018, Montréal, Canada.
[4]
Thomas Alsop. 2020. Global hourly enterprise server downtime cost 2019. https://www.statista.com/statistics/753938/
worldwide-enterprise-serverhourly-downtime- cost/ [Website accessed on 02-01-2021].
[5]
Backblaze. [n.d.]. Hard Drive Data and Stats. https://www.backblaze.com/b2/hard-drive-test-data.html [Website
accessed on 02-01-2021].
[6]
Amitabha Banerjee, Chien-Chia Chen, Chien-Chun Hung, Xiaobo Huang, Yifan Wang, and Razvan Chevesaran.
2020. Challenges and Experiences with MLOps for Performance Diagnostics in Hybrid-Cloud Enterprise Software
Deployments. In 2020 USENIX Conference on Operational Machine Learning, OpML 2020, July 28 - August 7, 2020.
[7]
Abdul Ali Bangash, Hareem Sahar, Abram Hindle, and Karim Ali. 2020. On the time-based conclusion stability of
cross-project defect prediction models. Empirical Software Engineering 25, 6 (2020), 5047–5083.
[8]
Chetan Bansal, Sundararajan Renganathan, Ashima Asudani, Olivier Midy, and Mathru Janakiraman. 2020. DeCaf: di-
agnosing and triaging performance issues in large-scale cloud services. In ICSE-SEIP 2020: 42nd International Conference
on Software Engineering, Software Engineering in Practice, Seoul, South Korea, 27 June - 19 July, 2020.
[9]
James Bergstra and Yoshua Bengio. 2012. Random Search for Hyper-Parameter Optimization. J. Mach. Learn. Res.
(2012).
[10]
Viv Bewick, Liz Cheek, and Jonathan Ball. 2004. Statistics review 13: receiver operating characteristic curves. Critical
care (2004).
[11]
Mirela Madalina Botezatu, Ioana Giurgiu, Jasmina Bogojeska, and Dorothea Wiesmann. 2016. Predicting Disk
Replacement towards Reliable Data Centers. In Proceedings of the 22nd ACM SIGKDD International Conference on
Knowledge Discovery and Data Mining, San Francisco, CA, USA, August 13-17, 2016.
[12]
Kendrick Boyd, Vitor Santos Costa, Jesse Davis, and C David Page. 2012. Unachievable region in precision-recall space
and its eect on empirical evaluation. In Proceedings of the 29th International Conference on Machine Learning.
[13] Johan Bring. 1995. Variable importance by partitioning R 2. Quality and Quantity (1995).
[14]
Oana-Maria Camburu, Eleonora Giunchiglia, Jakob Foerster, Thomas Lukasiewicz, and Phil Blunsom. 2019. Can
I Trust the Explainer? Verifying Post-hoc Explanatory Methods. CoRR abs/1910.02065 (2019). arXiv:1910.02065
http://arxiv.org/abs/1910.02065
[15]
Boyuan Chen and Zhen Ming Jack Jiang. 2019. Extracting and studying the Logging-Code-Issue-Introducing changes
in Java-based large-scale open source software systems. Empirical Software Engineering (2019).
[16]
Junjie Chen, Xiaoting He, Qingwei Lin, Hongyu Zhang, Dan Hao, Feng Gao, Zhangwei Xu, Yingnong Dang, and
Dongmei Zhang. 2019. Continuous Incident Triage for Large-Scale Online Service Systems. In 34th IEEE/ACM
International Conference on Automated Software Engineering, ASE 2019, San Diego, CA, USA, November 11-15, 2019.
[17]
Junjie Chen, Shu Zhang, Xiaoting He, Qingwei Lin, Hongyu Zhang, Dan Hao, Yu Kang, Feng Gao, Zhangwei Xu,
Yingnong Dang, and Dongmei Zhang. 2020. How Incidental are the Incidents? Characterizing and Prioritizing Incidents
for Large-Scale Online Service Systems. In 35th IEEE/ACM International Conference on Automated Software Engineering,
ASE 2020, Melbourne, Australia, September 21-25, 2020.
[18]
Yujun Chen, Xian Yang, Hang Dong, Xiaoting He, Hongyu Zhang, Qingwei Lin, Junjie Chen, Pu Zhao, Yu Kang, Feng
Gao, Zhangwei Xu, and Dongmei Zhang. 2020. Identifying linked incidents in large-scale online service systems. In
ESEC/FSE ’20: 28th ACM Joint European Software Engineering Conference and Symposium on the Foundations of Software
Engineering, Virtual Event, USA, November 8-13, 2020.
[19]
Yujun Chen, Xian Yang, Qingwei Lin, Hongyu Zhang, Feng Gao, Zhangwei Xu, Yingnong Dang, Dongmei Zhang,
Hang Dong, Yong Xu, Hao Li, and Yu Kang. 2019. Outage Prediction and Diagnosis for Cloud Service Systems. In The
World Wide Web Conference, WWW 2019, San Francisco, CA, USA, May 13-17, 2019. 2659–2665.
[20]
Zhuangbin Chen, Yu Kang, Liqun Li, Xu Zhang, Hongyu Zhang, Hui Xu, Yangfan Zhou, Li Yang, Jerey Sun, Zhangwei
Xu, et al
.
2020. Towards intelligent incident management: why we need it and how we make it. In Proceedings of the
28th ACM Joint Meeting on European Software Engineering Conference and Symposium on the Foundations of Software
Engineering. 1487–1497.
ACM Trans. Softw. Eng. Methodol., Vol. , No. , Article . Publication date: August 2021.
36 Yingzhe Lyu, Gopi Krishnan Rajbahadur∗, Dayi Lin∗, Boyuan Chen∗, and Zhen Ming (Jack) Jiang
[21]
Yingnong Dang, Qingwei Lin, and Peng Huang. 2019. AIOps: real-world challenges and research innovations. In
Proceedings of the 41st International Conference on Software Engineering: Companion Proceedings, ICSE 2019, Montreal,
QC, Canada, May 25-31, 2019.
[22]
Nosayba El-Sayed, Hongyu Zhu, and Bianca Schroeder. 2017. Learning from Failure Across Multiple Clusters: A
Trace-Driven Approach to Understanding, Predicting, and Mitigating Job Terminations. In 2017 IEEE 37th International
Conference on Distributed Computing Systems (ICDCS).
[23]
Ming Fan, Wenying Wei, Xiaofei Xie, Yang Liu, Xiaohong Guan, and Ting Liu. 2021. Can We Trust Your Explanations?
Sanity Checks for Interpreters in Android Malware Analysis. IEEE Trans. Inf. Forensics Secur. (2021).
[24]
João Gama, Pedro Medas, Gladys Castillo, and Pedro Pereira Rodrigues. 2004. Learning with Drift Detection. In
Advances in Articial Intelligence - SBIA 2004, 17th Brazilian Symposium on Articial Intelligence, São Luis, Maranhão,
Brazil, September 29 - October 1, 2004, Proceedings.
[25]
Timnit Gebru, Jamie Morgenstern, Briana Vecchione, Jennifer Wortman Vaughan, Hanna M. Wallach, Hal Daumé III,
and Kate Crawford. 2018. Datasheets for Datasets. CoRR abs/1803.09010 (2018). arXiv:1803.09010 http://arxiv.org/abs/
1803.09010
[26]
Odd Erik Gundersen and Sigbjørn Kjensmo. 2018. State of the Art: Reproducibility in Articial Intelligence. In
Proceedings of the Thirty-Second AAAI Conference on Articial Intelligence, (AAAI-18), the 30th innovative Applications of
Articial Intelligence (IAAI-18), and the 8th AAAI Symposium on Educational Advances in Articial Intelligence (EAAI-18),
New Orleans, Louisiana, USA, February 2-7, 2018.
[27]
Wenbo Guo, Dongliang Mu, Jun Xu, Purui Su, Gang Wang, and Xinyu Xing. 2018. Lemna: Explaining deep learning
based security applications. In proceedings of the ACM SIGSAC Conference on Computer and Communications Security.
[28]
Maayan Harel, Shie Mannor, Ran El-Yaniv, and Koby Crammer. 2014. Concept Drift Detection Through Resampling. In
Proceedings of the 31th International Conference on Machine Learning, ICML 2014, Beijing, China, 21-26 June 2014.
[29]
Shilin He, Qingwei Lin, Jian-Guang Lou, Hongyu Zhang, Michael R. Lyu, and Dongmei Zhang. 2018. Identifying
impactful service system problems via log analysis. In Proceedings of the 2018 ACM Joint Meeting on European Software
Engineering Conference and Symposium on the Foundations of Software Engineering, ESEC/SIGSOFT FSE 2018, Lake Buena
Vista, FL, USA, November 04-09, 2018.
[30]
Reyhane Askari Hemmat and Abdelhakim Had. 2016. SLA Violation Prediction In Cloud Computing: A Machine
Learning Perspective. CoRR (2016).
[31] Matthew Hutson. 2018. Articial intelligence faces reproducibility crisis. Science (2018).
[32]
IDG. 2020. 2020 Cloud Computing Study. https://www.idg.com/tools-for- marketers/2020-cloud- computing-study/
[Website accessed on 02-01-2021].
[33]
Ali Imran Jehangiri, Ramin Yahyapour, Philipp Wieder, Edwin Yaqub, and Kuan Lu. 2014. Diagnosing Cloud Performance
Anomalies Using Large Time Series Dataset Analysis. In 2014 IEEE 7th International Conference on Cloud Computing,
Anchorage, AK, USA, June 27 - July 2, 2014.
[34] George Jenks. 1967. The Data Model Concept in Statistical Mapping. In International Yearbook of Cartography.
[35]
Jiajun Jiang, Weihai Lu, Junjie Chen, Qingwei Lin, Pu Zhao, Yu Kang, Hongyu Zhang, Yingfei Xiong, Feng Gao,
Zhangwei Xu, Yingnong Dang, and Dongmei Zhang. 2020. How to mitigate the incident? an eective troubleshooting
guide recommendation technique for online service systems. In ESEC/FSE ’20: 28th ACM Joint European Software
Engineering Conference and Symposium on the Foundations of Software Engineering, Virtual Event, USA, November 8-13,
2020.
[36]
Jirayus Jiarpakdee, Chakkrit Tantithamthavorn, Hoa Khanh Dam, and John Grundy. 2020. An empirical study of
model-agnostic techniques for defect prediction models. IEEE Transactions on Software Engineering (2020).
[37]
Jirayus Jiarpakdee, Chakkrit Tantithamthavorn, and Christoph Treude. 2020. The impact of automated feature selection
techniques on the interpretation of defect models. Empirical Software Engineering 25, 5 (2020), 3590–3638.
[38]
Rahul Krishna and Tim Menzies. 2018. Bellwethers: A baseline method for transfer learning. IEEE Transactions on
Software Engineering (2018).
[39]
William H Kruskal and W Allen Wallis. 1952. Use of ranks in one-criterion variance analysis. Journal of the American
statistical Association 47, 260 (1952), 583–621.
[40]
Jing Li, Xinpu Ji, Yuhan Jia, Bingpeng Zhu, Gang Wang, Zhongwei Li, and Xiaoguang Liu. 2014. Hard Drive Failure
Prediction Using Classication and Regression Trees. In 44th Annual IEEE/IFIP International Conference on Dependable
Systems and Networks, DSN 2014, Atlanta, GA, USA, June 23-26, 2014.
[41]
Jing Li, Rebecca J. Stones, Gang Wang, Xiaoguang Liu, Zhongwei Li, and Ming Xu. 2017. Hard drive failure prediction
using Decision Trees. Reliab. Eng. Syst. Saf. (2017).
[42]
Yangguang Li, Zhen Ming (Jack) Jiang, Heng Li, Ahmed E. Hassan, Cheng He, Ruirui Huang, Zhengda Zeng, Mian
Wang, and Pinan Chen. 2020. Predicting Node Failures in an Ultra-Large-Scale Cloud Computing Platform: An AIOps
Solution. ACM Trans. Softw. Eng. Methodol. (2020).
ACM Trans. Softw. Eng. Methodol., Vol. , No. , Article . Publication date: August 2021.
Towards a consistent interpretation of AIOps models 37
[43]
Cynthia C. S. Liem and Annibale Panichella. 2020. Run, Forest, Run? On Randomization and Reproducibility in
Predictive Software Engineering. arXiv:2012.08387 (2020).
[44]
Meng-Hui Lim, Jian-Guang Lou, Hongyu Zhang, Qiang Fu, Andrew Beng Jin Teoh, Qingwei Lin, Rui Ding, and
Dongmei Zhang. 2014. Identifying Recurrent and Unknown Performance Issues. In 2014 IEEE International Conference
on Data Mining, ICDM 2014, Shenzhen, China, December 14-17, 2014. 320–329.
[45]
Dayi Lin, Cor-Paul Bezemer, and Ahmed E Hassan. 2018. An empirical study of early access games on the Steam
platform. Empirical Software Engineering (2018).
[46]
Dayi Lin, Cor-Paul Bezemer, and Ahmed E Hassan. 2019. Identifying gameplay videos that exhibit bugs in computer
games. Empirical Software Engineering (2019).
[47]
Qingwei Lin, Ken Hsieh, Yingnong Dang, Hongyu Zhang, Kaixin Sui, Yong Xu, Jian-Guang Lou, Chenggang Li, Youjiang
Wu, Randolph Yao, Murali Chintalapati, and Dongmei Zhang. 2018. Predicting Node failure in cloud service systems.
In Proceedings of the 2018 ACM Joint Meeting on European Software Engineering Conference and Symposium on the
Foundations of Software Engineering, ESEC/SIGSOFT FSE 2018, Lake Buena Vista, FL, USA, November 04-09, 2018.
[48]
Qingwei Lin, Jian-Guang Lou, Hongyu Zhang, and Dongmei Zhang. 2016. iDice: problem identication for emerging
issues. In Proceedings of the 38th International Conference on Software Engineering, ICSE 2016, Austin, TX, USA, May
14-22, 2016.
[49]
Zachary C Lipton. 2018. The Mythos of Model Interpretability: In machine learning, the concept of interpretability is
both important and slippery. Queue (2018).
[50]
Jian-Guang Lou, Qingwei Lin, Rui Ding, Qiang Fu, Dongmei Zhang, and Tao Xie. 2013. Software analytics for incident
management of online services: An experience report. In 2013 28th IEEE/ACM International Conference on Automated
Software Engineering, ASE 2013, Silicon Valley, CA, USA, November 11-15, 2013.
[51]
Jian-Guang Lou, Qingwei Lin, Rui Ding, Qiang Fu, Dongmei Zhang, and Tao Xie. 2017. Experience report on applying
software analytics in incident management of online service. Autom. Softw. Eng. (2017).
[52]
Scott M Lundberg and Su-In Lee. 2017. A unied approach to interpreting model predictions. In Proceedings of the 31st
International Conference on Neural Information Processing Systems.
[53]
Chen Luo, Jian-Guang Lou, Qingwei Lin, Qiang Fu, Rui Ding, Dongmei Zhang, and Zhe Wang. 2014. Correlating events
with time series for incident diagnosis. In The 20th ACM SIGKDD International Conference on Knowledge Discovery and
Data Mining, KDD ’14, New York, NY, USA - August 24 - 27, 2014.
[54]
Yingzhe Lyu. 2021. Replication Package for An Empirical Study of the Impact of Data Splitting Decisions on the
Performance of AIOps Solutions. https://github.com/SAILResearch/suppmaterial-19- yingzhe-aiops_data_splitting
[Website accessed on 06-13-2021].
[55]
Yingzhe Lyu, Heng Li, Mohammed Sayagh, Zhenming (Jack) Jiang, and Ahmed E. Hassan. 2021. An Empirical Study
of the Impact of Data Splitting Decisions on the Performance of AIOps Solutions. ACM Trans. Softw. Eng. Methodol.
(2021).
[56]
Farzaneh Mahdisoltani, Ioan A. Stefanovici, and Bianca Schroeder. 2017. Proactive error prediction to improve storage
system reliability. In 2017 USENIX Annual Technical Conference, USENIX ATC 2017, Santa Clara, CA, USA, July 12-14,
2017. 391–402.
[57] Tim Menzies. 2019. The ve laws of SE for AI. IEEE Software 37, 1 (2019), 81–85.
[58]
Kit Merker. 2020. Council Post: How To Optimize Gross Margins For Digital Services. https:
//www.forbes.com/sites/forbesbusinessdevelopmentcouncil/2020/12/03/how-to-optimize- gross-margins-for-
digital-services/?sh=13d62bb130c3 [Website accessed on 02-01-2021].
[59] Rupert G. Miller. 2012. Simultaneous Statistical Inference. Springer Science & Business Media.
[60]
Margaret Mitchell, Simone Wu, Andrew Zaldivar, Parker Barnes, Lucy Vasserman, Ben Hutchinson, Elena Spitzer,
Inioluwa Deborah Raji, and Timnit Gebru. 2019. Model Cards for Model Reporting. In Proceedings of the Conference on
Fairness, Accountability, and Transparency, FAT* 2019, Atlanta, GA, USA, January 29-31, 2019.
[61] Christoph Molnar. 2019. Interpretable Machine Learning. https://christophm.github.io/interpretable-ml-book/.
[62]
Sasho Nedelkoski, Jorge S. Cardoso, and Odej Kao. 2019. Anomaly Detection and Classication using Distributed
Tracing and Deep Learning. In 19th IEEE/ACM International Symposium on Cluster, Cloud and Grid Computing, CCGRID
2019, Larnaca, Cyprus, May 14-17, 2019.
[63]
Dino Pedreschi, Fosca Giannotti, Riccardo Guidotti, Anna Monreale, Salvatore Ruggieri, and Franco Turini. 2019.
Meaningful explanations of Black Box AI decision systems. In Proceedings of the AAAI Conference on Articial
Intelligence.
[64]
Hung Viet Pham, Shangshu Qian, Jiannan Wang, Thibaud Lutellier, Jonathan Rosenthal, Lin Tan, Yaoliang Yu, and
Nachiappan Nagappan. 2020. Problems and Opportunities in Training Deep Learning Software Systems: An Analysis
of Variance. In 35th IEEE/ACM International Conference on Automated Software Engineering (ASE).
[65]
Joelle Pineau, Philippe Vincent-Lamarre, Koustuv Sinha, Vincent Larivière, Alina Beygelzimer, Florence d’Alché Buc,
Emily Fox, and Hugo Larochelle. 2020. Improving Reproducibility in Machine Learning Research (A Report from the
ACM Trans. Softw. Eng. Methodol., Vol. , No. , Article . Publication date: August 2021.
38 Yingzhe Lyu, Gopi Krishnan Rajbahadur∗, Dayi Lin∗, Boyuan Chen∗, and Zhen Ming (Jack) Jiang
NeurIPS 2019 Reproducibility Program). arXiv:2003.12206 [cs.LG]
[66]
Gopi Krishnan Rajbahadur, Shaowei Wang, Gustavo A. Oliva, Yasutaka Kamei, and Ahmed E. Hassan. 2021. The impact
of feature importance methods on the interpretation of defect classiers. In IEEE Transactions on Software Engineering.
[67]
Baishakhi Ray, Vincent Hellendoorn, Saheel Godhane, Zhaopeng Tu, Alberto Bacchelli, and Premkumar Devanbu.
2016. On the" naturalness" of buggy code. In 2016 IEEE/ACM 38th International Conference on Software Engineering
(ICSE). IEEE, 428–439.
[68]
Charles Reiss, John Wilkes, and Joseph L Hellerstein. 2011. Google cluster-usage traces: format+ schema. Google Inc.,
White Paper (2011).
[69]
Marco Túlio Ribeiro, Sameer Singh, and Carlos Guestrin. 2016. "Why Should I Trust You?": Explaining the Predictions
of Any Classier. In Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data
Mining, San Francisco, CA, USA, August 13-17, 2016.
[70]
Marco Tulio Ribeiro, Sameer Singh, and Carlos Guestrin. 2018. Anchors: High-precision model-agnostic explanations.
In Proceedings of the AAAI Conference on Articial Intelligence.
[71]
Andrea Rosà, Lydia Y. Chen, and Walter Binder. 2015. Catching failures of failures at big-data clusters: A two-level
neural network approach. In 23rd IEEE International Symposium on Quality of Service, IWQoS 2015, Portland, OR, USA,
June 15-16, 2015.
[72]
Cynthia Rudin. 2019. Stop explaining black box machine learning models for high stakes decisions and use interpretable
models instead. Nature Machine Intelligence (2019).
[73]
W Nick Street and YongSeog Kim. 2001. A streaming ensemble algorithm (SEA) for large-scale classication. In
Proceedings of the seventh ACM SIGKDD international conference on Knowledge discovery and data mining.
[74]
Chakkrit Tantithamthavorn and Ahmed E. Hassan. 2018. An Experience Report on Defect Modelling in Practice:
Pitfalls and Challenges. In In Proceedings of the International Conference on Software Engineering: Software Engineering
in Practice Track (ICSE-SEIP’18).
[75]
Chakkrit Tantithamthavorn, Ahmed E Hassan, and Kenichi Matsumoto.2018. The impact of class rebalancing techniques
on the performance and interpretation of defect prediction models. IEEE Transactions on Software Engineering 46, 11
(2018), 1200–1219.
[76]
Chakkrit Tantithamthavorn, Shane McIntosh, Ahmed E Hassan, and Kenichi Matsumoto. 2018. The impact of automated
parameter optimization on defect prediction models. IEEE Transactions on Software Engineering 45, 7 (2018), 683–711.
[77]
D Roland Thomas, Edward Hughes, and Bruno D Zumbo. 1998. On variable importance in linear regression. Social
Indicators Research (1998).
[78] Robert L. Thorndike. 1953. Who belongs in the family. Psychometrika (1953).
[79]
Burak Turhan. 2012. On the dataset shift problem in software engineering prediction models. Empirical Software
Engineering (2012).
[80]
Haixun Wang, Wei Fan, Philip S Yu, and Jiawei Han. 2003. Mining concept-drifting data streams using ensemble
classiers. In Proceedings of the ninth ACM SIGKDD international conference on Knowledge discovery and data mining.
[81]
Alexander Warnecke, Daniel Arp, Christian Wressnegger, and Konrad Rieck. 2020. Evaluating Explanation Methods for
Deep Learning in Security. In IEEE European Symposium on Security and Privacy, EuroS&P 2020, Genoa, Italy, September
7-11, 2020.
[82]
Yong Xu, Kaixin Sui, Randolph Yao, Hongyu Zhang, Qingwei Lin, Yingnong Dang, Peng Li, Keceng Jiang, Wenchi
Zhang, Jian-Guang Lou, Murali Chintalapati, and Dongmei Zhang. 2018. Improving Service Availability of Cloud
Systems by Predicting Disk Error. In 2018 USENIX Annual Technical Conference, USENIX ATC 2018, Boston, MA, USA,
July 11-13, 2018.
[83]
Ji Xue, Robert Birke, Lydia Y. Chen, and Evgenia Smirni. 2016. Managing Data Center Tickets: Prediction and Active
Sizing. In 46th Annual IEEE/IFIP International Conference on Dependable Systems and Networks, DSN 2016, Toulouse,
France, June 28 - July 1, 2016.
[84]
Ji Xue, Robert Birke, Lydia Y. Chen, and Evgenia Smirni. 2018. Spatial-Temporal Prediction Models for Active Ticket
Managing in Data Centers. IEEE Trans. Netw. Serv. Manag. (2018).
[85]
Chih-Kuan Yeh, Cheng-Yu Hsieh, Arun Sai Suggala, David I. Inouye, and Pradeep Ravikumar. 2019. On the (In)delity
and Sensitivity of Explanations. In Advances in Neural Information Processing Systems 32: Annual Conference on Neural
Information Processing Systems 2019, NeurIPS 2019, December 8-14, 2019, Vancouver, BC, Canada.
[86]
Xuezhi Zeng, Saurabh Garg, Mutaz Barika, Sanat Kumar Bista, Deepak Puthal, Albert Y. Zomaya, and Rajiv Ranjan.
2021. Detection of SLA Violation for Big Data Analytics Applications in Cloud. IEEE Trans. Computers (2021).
[87]
Steve Zhang, Ira Cohen, Moisés Goldszmidt, Julie Symons, and Armando Fox. 2005. Ensembles of Models for Automated
Diagnosis of System Performance Problems. In 2005 International Conference on Dependable Systems and Networks
(DSN 2005), 28 June - 1 July 2005, Yokohama, Japan, Proceedings.
[88]
Nengwen Zhao, Junjie Chen, Zhou Wang, Xiao Peng, Gang Wang, Yong Wu, Fang Zhou, Zhen Feng, Xiaohui Nie,
Wenchi Zhang, Kaixin Sui, and Dan Pei. 2020. Real-time incident prediction for online service systems. In ESEC/FSE ’20:
ACM Trans. Softw. Eng. Methodol., Vol. , No. , Article . Publication date: August 2021.
Towards a consistent interpretation of AIOps models 39
28th ACM Joint European Software Engineering Conference and Symposium on the Foundations of Software Engineering,
Virtual Event, USA, November 8-13, 2020.
ACM Trans. Softw. Eng. Methodol., Vol. , No. , Article . Publication date: August 2021.
