Conference PaperPDF Available

Abstract and Figures

A fundamental assumption of Business Process Management (BPM) is that redesign delivers new and improved versions of business processes. This assumption, however, does not necessarily hold, and required compensatory action may be delayed until a new round in the BPM life-cycle completes. Current approaches to process redesign face this problem in one way or another, which makes rapid process improvement a central research problem of BPM today. In this paper, we address this problem by integrating concepts from process execution with ideas from DevOps. More specifically, we develop a technique called AB-BPM that offers AB testing for process versions with immediate feedback at runtime. We implemented this technique in such a way that two versions (A and B) are operational in parallel and any new process instance is routed to one of them. The routing decision is made at runtime on the basis of the achieved results for the registered performance metrics of each version. AB-BPM provides for ultimate convergence towards the best performing version, no matter if it is the old or the new version. We demonstrate the efficacy of our technique by conducting an extensive evaluation based on both synthetic and real-life data.
Content may be subject to copyright.
AB-BPM: Performance-driven Instance Routing
for Business Process Improvement
Suhrid Satyal1,2, Ingo Weber1,2, Hye-young Paik2,
Claudio Di Ciccio3, and Jan Mendling3
1Data61, CSIRO, Sydney, Australia
{suhrid.satyal,ingo.weber}@data61.csiro.au
2University of New South Wales, Sydney, Australia
hpaik@cse.unsw.edu.au
3Vienna University of Economics and Business, Vienna, Austria
{claudio.di.ciccio,jan.mendling}@wu.ac.at
Abstract. A fundamental assumption of Business Process Management
(BPM) is that redesign delivers new and improved versions of business
processes. This assumption, however, does not necessarily hold, and re-
quired compensatory action may be delayed until a new round in the
BPM life-cycle completes. Current approaches to process redesign face
this problem in one way or another, which makes rapid process improve-
ment a central research problem of BPM today. In this paper, we address
this problem by integrating concepts from process execution with ideas
from DevOps. More specifically, we develop a technique called AB-BPM
that offers AB testing for process versions with immediate feedback at
runtime. We implemented this technique in such a way that two versions
(A and B) are operational in parallel and any new process instance is
routed to one of them. The routing decision is made at runtime on the
basis of the achieved results for the registered performance metrics of
each version. AB-BPM provides for ultimate convergence towards the
best performing version, no matter if it is the old or the new version. We
demonstrate the efficacy of our technique by conducting an extensive
evaluation based on both synthetic and real-life data.
Keywords: Business process management, DevOps, AB Testing, Pro-
cess performance indicators
1 Introduction
Various lifecycle approaches to Business Process Management (BPM) have a
common assumption that a process is incrementally improved in the redesign
phase [9, Ch.1]. While this assumption is hardly questioned in BPM research,
there is evidence from the field of AB testing that improvement concepts often
do not lead to actual improvements. For instance, work on business improvement
ideas found that 75 percent did not lead to improvement: half of them had no
impact while approximately a quarter turned out to be even harmful [12]. The
results are comparable to a study of the Microsoft website, in which only one
Pre-print copy of the manuscript published by Springer
(available at link.springer.com)
identified by doi:10.1007/978-3-319-65000-5_7
2 S. Satyal, I. Weber, H. Paik, C. Di Ciccio, J. Mendling
third of the ideas observed had a positive impact, while the remaining had no or
negative impact [15]. The same study also observed that customer preferences
were difficult to anticipate before deployment, and that customer research did
not predict customer behaviour accurately.
If incremental process improvement can only be achieved in a fraction of
the cases, there is a need to rapidly validate the assumed benefits. Unfortu-
nately, there are currently two major challenges for such an immediate valida-
tion. The first one is methodological. Classical BPM lifecycle approaches build
on a labour-intensive analysis of the current process, which leads to the de-
ployment of a redesigned version. This new version is monitored in operation,
and if it does not meet performance objectives, it is made subject to analysis
again. All this takes time. The second challenge is architectural. Contemporary
Business Process Management Systems (BPMSs) enable quick deployment of
process improvements, but they do not offer support for validating improvement
assumptions. A performance comparison between the old and the new version
may be biased since contextual factors might have changed at the same time.
How a rapid validation of improvement assumptions can be integrated in the
BPM lifecycle and in BPMSs is an open research question.
In this paper, we address this question by integrating business process execu-
tion concepts with ideas from DevOps. More specifically, we develop a technique
called AB-BPM that offers AB testing for redesigned processes with immediate
feedback at runtime. AB testing in DevOps compares two versions of a deployed
product (e.g., a Web page) by observing users’ responses to versions A/B, and
determines which one performs better [8]. We implemented this technique in
such a way that two versions (A and B) of a process are operational in par-
allel and any new process instance is routed to one of them. Through a series
of experiments and observations, we have developed an instance routing algo-
rithm, LTAvgR, which is adapted to the context of executing business processes.
The routing decision is guided by the observed results for registered performance
metrics of each version at runtime. The technique has been evaluated extensively
on both synthetic and real-life data. The results showed that AB-BPM provides
for ultimate convergence towards the best performing version.
The remainder of this paper starts with a discussion of the background and
related work in Section 2.Section 3 describes the framework and algorithms for
performing AB tests. In Section 4, we evaluate our approach on two use cases.
In Section 5, we discuss the results, limitations, and validity of our approach,
and finally draw conclusions in Section 6.
2 Background and Related Work
Business Process Management Systems (BPMSs) allow for a rapid deployment
of process improvements into operation. However, there is currently no support
to test the often implicit assumption that a modification of a process actually
represents an improvement. One anecdote of a leading European bank (EB) il-
lustrates this problem. The EB improved their loan approval process by cutting
Pre-print copy of the manuscript published by Springer
(available at link.springer.com)
identified by doi:10.1007/978-3-319-65000-5_7
AB-BPM: Performance-driven Instance Routing 3
its turnaround time down from one week to a few hours. What happened though
was a steep decline in customer satisfaction: customers with a negative notice
would complain that their application might have been declined unjustifiably;
customers with a positive notice would inquire whether their application had
been checked with due diligence. This anecdote emphasizes the need to carefully
test improvement hypotheses in practice since customers and process partici-
pants might not act in a way that can be predicted deterministically.
Given the current architecture of BPMSs, it is not possible to conduct a
fair comparison between the old and the new version of a process since they
are not operational at the same point of time. That means, doing a post-hoc
analysis of data generated from the old process being operational in time interval
rtpn´1q, tpnqs and the new process running from rtpnq, tpn`1qs is biased towards
the respective conditions of each time interval.
The need for continuous improvement is rarely disputed, but it should be
complemented with the motto “test fairly” and “fail fast”. This motto entails
reducing the time between inception and deployment of new versions, and mang-
ing business risks brought forth by deployment of the new process versions.
From the above analysis, we derive the following three requirements:
R1 Rapidly validate the improvement assumption: a proposed improvement
should be tested within a short time frame after its introduction.
R2 Ensure a fair comparison: the environment in which a comparison is con-
ducted should minimize bias, and avoid the time bias discussed above.
R3 Enable rapid adjustments on process model level: the benefits of a solution
should be suited to process models and their specific characteristics.
Concepts from DevOps [3], which aim to bring software development (Dev)
and operations (Ops) closer together, may help to address this problem. One
DevOps practice is live testing, where new versions of the software are tested
in production with actual users of the system. The most popular form of live
testing is AB testing, where two versions (Aand B) are deployed side by side and
both receive a share of the production workload while being monitored closely.
The monitoring data is then used to draw conclusions about the effectiveness of
one version over the other, for instance in the form of increased revenue from
higher click-through rates.
So far, AB testing has been used for micro changes of websites, like changing
the color of a button [8,15]. The effectiveness of this technique is surveyed by
Kohavi et al. for the user interfaces of web applications [16,15]. In this paper,
we adopt the idea of AB testing on the process level in order to address R1-
R3. Our technique is called AB-BPM. In the following, we discuss in how far
previous BPM research is related to R1- R3. We distinguish methodological and
technical approaches to process improvement.
Methodological approaches include business process re-engineering and the
BPM lifecycle. Business Process Re-Engineering (BPR) offers a methodology for
selecting and redesigning a process to improve efficiency, often exploiting IT to
support the changes [11]. BPR promotes radical changes to the processes. An
explicit perspective for testing re-engineered processes is missing.
Pre-print copy of the manuscript published by Springer
(available at link.springer.com)
identified by doi:10.1007/978-3-319-65000-5_7
4 S. Satyal, I. Weber, H. Paik, C. Di Ciccio, J. Mendling
Table 1: Mapping existing works to the requirements
Approach R1 R2 R3
Process Re-engineering (Hammer/Champy) [11] - - -
Process Improvement [14] - - -
Process Lifecycle [9] +/- - -
Statistical Process Control (SPC) [13] + - -
Complex Event Processing (CEP) [24,25] - - +
AB Testing, see e.g., [3,15] + + -
AB-BPM (this work) + + +
Kettinger et. al. summarize methods for process improvement project [14].
Lifecycle models like the one described by Dumas et. al. [9] propose a more in-
cremental improvement of processes with periodic controlling and revisions. Our
research complements this stream of research by providing techniques to exper-
iment with process improvement hypotheses and perform statistical evaluation
on them. We assume that a redesigned process is made available, so that such
experimentation and analysis can be done. We envision that our approach can
be used in conjunction with works that automatically generate process versions
just as [4]. Other techniques include root-cause analysis [19,9], e.g. by the help of
cause-effect diagrams [10,9], and the consideration of best-practises [20,2]. How-
ever, evaluation of effectiveness requires the involvement of process analysts.
Technical approaches focus on monitoring processes at runtime with a focus
on specific performance metrics. Concepts based on Statistical Process Control
(SPC) [13], Complex Event Processing (CEP) [24,25], and predictive analytics
[6] have been proposed and adapted for monitoring business processes. How-
ever, these monitoring techniques have not been used to carry out controlled
experiments. In our work, the monitoring is performed by the instance router
by observing a Process Performance Indicator (PPI) like satisfaction ratings ob-
tained from end users. Based on the chosen PPI, the instance router dynamically
adjusts the request distribution rates.
Our research addresses the gap of an explicit testing of improvement hy-
potheses in BPM-related research and the lack of an explicit consideration of
business processes in the works on AB testing. In the following, we devise our
AB-BPM approach so that it meets requirements R1-R3.
3 Approach and Architecture
In this section, we present our approach, starting with mapping the instance
routing problem to algorithms from the literature. Based on a small experiment,
we choose one algorithm and adapt it to the context of business processes. Then
we present our high-level framework, architecture, and implementation.
3.1 Instance Routing – a Multi-Armed Bandit Problem
In order to integrate concepts of process execution with AB testing, we have
to discuss how new instances are assigned to a specific version of the process.
Pre-print copy of the manuscript published by Springer
(available at link.springer.com)
identified by doi:10.1007/978-3-319-65000-5_7
AB-BPM: Performance-driven Instance Routing 5
Therefore, we need an instance router that distributes requests to versions in such
a way that any relevant Process Performance Indicator (PPI) is maximized. The
instance router also needs to deal effectively with the issue that processes can
be long-running, and that PPI measurements can be delayed.
The PPI maximization can be mapped to the so-called multi-armed bandit
problem [5,1]. The multi-armed bandit problem models a hypothetical experi-
ment where, given some slot machines with different payoff probability distribu-
tions, a gambler has to decide on which machines to play. The objective of the
gambler is to maximize the total payoff during a sequence of plays. Since the
gamblers are unaware of the payoff distribution, they can approach the plays
with two strategies: exploring the payoffs by pulling different arms on the ma-
chines or exploiting the current knowledge by pulling arms that are known to
give good payoffs. The exploration strategy builds knowledge about the pay-
offs, and the exploitation strategy accumulates the payoffs. Multi-armed bandit
algorithms aim to find a balance between these strategies. If the performance
is affected by some context, this can be seen as the so-called contextual multi-
armed bandit problem, where the gambler sees context (typically represented as
a multi-dimensional feature vector) associated with the current iteration before
making the choice.
We model the routing algorithm as a multi-armed bandit problem by repre-
senting the process versions as the “arms”, and the PPI as “payoffs/rewards”.
The objective of the instance router is to find a good tradeoff between explo-
ration and exploitation, possibly based on the context. To learn the performance
of a version in exploration, it sends some of the process instantiation requests
to either version. Based on the instance router’s experience, it can exploit its
knowledge to send more or even all request to the better-performing version. The
reward for routing algorithm can be designed to use a PPI like user satisfaction.
3.2 Instance Routing Algorithms & Selection
The multi-armed bandit problem has been explored in related literature. LinUCB
[17] is a contextual multi-armed bandit algorithm that has been employed to
serve news articles to users with the objective to maximize the total number
of clicks. Tompson sampling [23] is one of the simplest approaches to address
multi-armed bandits. It is based on a Bayesian approach where arms are chosen
according to their probability of producing optimal rewards [7,23]. Thompson
sampling can be used to solve the contextual multi-armed bandit problem with
linear rewards [1]. In this paper, we chose these three algorithms as candidates
for process instance routing and investigate their effectiveness. We have selected
these algorithms based on their demonstrated benefits and simplicity. Other
algorithms, such as -greedy, -first, UCB, and EXP4 also address multi-armed
bandit problems [5].
Since the goal for the routing algorithms is to maximize the cumulative value
of the PPI, as preparatory work, we have experimented with different routing al-
gorithms with different configurations to find the best performing algorithm. We
have compared variations of Thompson sampling techniques [1,7,23], LinUCB[17]
Pre-print copy of the manuscript published by Springer
(available at link.springer.com)
identified by doi:10.1007/978-3-319-65000-5_7
6 S. Satyal, I. Weber, H. Paik, C. Di Ciccio, J. Mendling
and a baseline algorithm which uniformly distributes requests to process ver-
sions regardless of context and rewards. We have found that LinUCB produced
the highest cumulative satisfaction score throughout the experiments. Therefore
we use this algorithm in the following. Our architecture is flexible enough that
it can be easily replaced by other algorithms.
3.3 Adapting the Routing Algorithm to Business Processes
As discussed above, we chose LinUCB as our routing algorithm. However, we
observed that the algorithm can be derailed by process-specific circumstances,
such as the long delays before rewards. Long delays are inherent to long-running
processes, and not considered in AB testing solutions for Web applications, where
delays are measured in seconds or minutes. In contrast, the real-world data which
we use in the evaluation has one process instance with an overall duration of more
than 3 years.
This results in the following issue. Oftentimes overly long process completion
times correlate with problematic process instances, leading to negative rewards.
Thus, instances with short completion times can give a positive impression of
a process version early on. If the algorithm receives too many positive rewards
from one version during the early stages of the AB test, the algorithm is more
likely to see that version as preferable. Such an early determination can introduce
a bias in the evaluation. Thus, we need to ensure that the algorithm gets enough
samples from both versions.
We solve this issue by adopting the idea of a “warm-up” phase from Rein-
forcement Learning [21], during which we emphasize exploration over exploita-
tion. We sample the probability of exploration by using an exponential decay
function, acting as an injected perturbation that diminishes as the experiment
proceeds – the sample determines whether the algorithm follows LinUCB’s de-
cision or picks a version at random. We consider the “warm-up” as the exper-
imentation phase: after all instances started during the experimentation phase
are completed, no more rewards are collected and the instance router stabilizes.
Finally, the original LinUCB algorithm makes its decision based on the sum-
mation of past rewards. We found out during the experiments that this can also
deceive the algorithm. Therefore, we have modified the LinUCB algorithm to
make its reward estimates on the basis of the average of past rewards rather
than their sum. We term our adapted instance routing algorithm Long-term
average router (LTAvgR).
3.4 AB-BPM Framework, Architecture, and Implementation
Figure 1 shows the architecture of our AB testing framework. We designed the
architecture such that the two versions of the process model are deployed side by
side in the same execution engine. The instance router distributes the instance
creation requests as per its internal logic. An alternative design would be to run
two full copies of the entire application stack, one for each version, and using
the instance router to distribute the requests across the two stacks. However,
Pre-print copy of the manuscript published by Springer
(available at link.springer.com)
identified by doi:10.1007/978-3-319-65000-5_7
AB-BPM: Performance-driven Instance Routing 7
Clients
RESTful API
Process
engine
Implemented
service
Version A Version B
Database
Instance
router
Routing
algorithm
1. Instantiate 2. Instantiate
3. Interact
4. Update reward
Fig. 1: Application architecture
the multi-armed bandit algorithms can identify the superior version during the
experimentation and alter the allocation of requests to different versions. When
a version is clearly superior to the other, most of the requests are sent to the
superior version. In such scenarios, the application stack that hosts the inferior
version is underutilized. If we run both versions on the same stack, we can keep
utilization of the system high, no matter which version is superior.
Given this design choice, the process definitions, implementation, PPI collec-
tion, and the shared process execution engine are wrapped by a web application.
Process execution data are stored in a shared database. Process intantiation,
metrics, and other operations are exposed using RESTful APIs. Upon receiving
a request, the instance router instantiates a particular version and receives an
identifier. Identifiers of process instances for which rewards have not been ob-
served are stored in a queue. The instance router uses a polling mechanism in
parallel to retrieve PPI metrics from the server and update the rewards.
We implemented the architecture prototypically in Java and Python, in part
based on the Activiti BPMS. As outlined earlier, our framework is flexible in
the choice of the instance routing algorithm: we implemented and tested all
five variants discussed above, i.e., LTAvgR, LinUCB, Thompson-sampling with
and without linear rewards, and random uniform distribution. The experiments
reported in the following section are run with the presented implementation with
LTAvgR. We simulate the requests from users by replaying process logs.
4 Evaluation
In this section, we present the methodology and outcomes of our evaluation of
the proposed approach. We assess the AB-BPM framework and LTAvgR algo-
rithm first on synthetic data, where we have full control over the environment
and parameters. Then we test the approach on real-world data, taken from the
building permit process logs from five Dutch municipalities.
4.1 Evaluation on Synthetic Data
In this section, we demonstrate our approach using two example process versions
stemming from the domain of helicopter pilot licensing. Version A of the process
sequentially schedules the activities based on the cost of performing them. Based
Pre-print copy of the manuscript published by Springer
(available at link.springer.com)
identified by doi:10.1007/978-3-319-65000-5_7
8 S. Satyal, I. Weber, H. Paik, C. Di Ciccio, J. Mendling
Eligibility
Test
Medical
Exam
Practical
Theory
Test
License
Processing
Schedule
Schedule
passed
Schedule
Schedule
passed
passed
failed
failed
(a) Version A
Medical
Exam
Practical
Test
Theory
Test
License
Processing
Schedule
Eligibility
Test
passed
passed
passed
failed
failed
failed
failed
(b) Version B
Fig. 2: Process versions for the AB Testing experiment
Table 2: Cost model of the activities
Activity Cost Min. processing time Max. processing time
Schedule 25 1 day 1 day
Eligibility Test 190 1 day 3 days
Medical Exam 75 1 day 3 days
Theory Test 455 2 weeks 5 weeks
Practical Test 1145 1 week 2 weeks
License Processing 0 if rejected, 100 if approved Immediate Immediate
on the result (pass/fail), the process either schedules the next activity or termi-
nates the process. In this version, we expect that successful candidates will pay
more because of multiple scheduling costs. In contrast, version B of the process
schedules all such activities at the beginning, thus reducing the scheduling costs.
The processes are illustrated in Fig. 2. These processes have as a result the final
status of the license: either approved or rejected.
As PPI we simulate the user satisfaction, here calculated as a combination
of cost, completion time, and result of the process execution. Costs and process-
ing times of each task were derived from the Australian Civil Aviation Safety
Authority (CASA)4and helicopter hiring rates from an Australian flight school.
Table 2 shows the costs and processing times for both process versions. Table 3
shows how user satisfaction scores from 1 (lowest) to 5 (highest) are derived. The
basic rationale is, the shorter and cheaper, the better. The score ranges from 1
to 3 if the outcome is negative, and from 3 to 5 if the outcome is positive.
Experiment design. We have designed the AB testing experiments such that the
instance router receives requests with embedded contexts at the rate of 1 request
per second. For the execution, we scale each day to 1 second. In this experiment,
we introduce non-determinism in two ways: (i) adjusting success rates of an
4https://www.legislation.gov.au/Details/F2016C00882, Accessed: 03-01-2017
Pre-print copy of the manuscript published by Springer
(available at link.springer.com)
identified by doi:10.1007/978-3-319-65000-5_7
AB-BPM: Performance-driven Instance Routing 9
Table 3: User satisfaction model
Outcome Cost Duration Satisfaction
Approved
r0,1990s ď 5 weeks 5
p1990,8s ď 5 weeks 4
r0,8q ą 5 weeks 3
Rejected
r0,1890s ď 5 weeks 3
r0,1890s ą 5 weeks 2
p1890,8s ď 5 weeks 2
p1890,8q ą 5 weeks 1
Fig. 3: Requests routing in AB Tests
activity based on contextual information, and (ii) sampling processing times for
each activity using a probability distribution function.
Results. Fig. 3 shows the request distribution throughout the AB Tests. When
the experimentation or “warm-up” phase ends at approximately 1000 requests,
the router stops updating the reward estimates and chooses version B decisively.
A post-hoc analysis shows that the median user satisfaction was similar for
both versions. The distributions of user satisfaction scores differed significantly
(a Mann-Whitney test [18] resulted in U=54072, p-value ă10´6two-tailed,
nA=222, nB=778). The median delay of the reward was 22.3 seconds. Table 4
shows the differences between the two versions. Version B produces a better user
satisfaction in those cases where an application is approved. It is also faster in
all cases. However, the median cost of version A is lower than that of version B
when the applications are rejected.
We used an evaluation based on synthetic log data in order to investigate the
convergence behaviour of the implementation. We observe that our approach
leads to a rapid identification of the more rewarding process version, which re-
ceives an increasing share of traffic. This observation is instrumental with respect
to the requirements R1 - R3, which demand rapid validation, fair comparison
and rapid adjustment on the process level.
4.2 Evaluation on Real-World Data
To assess the applicability of our approach over real-world data, we have anal-
ysed the data stemming from the five logs in the BPI Challenge 20155, herein
identified as L1, . . . , L5. Those logs contain the execution data of the building
permit issuing processes in five Dutch municipalities. The processes behind each
log reportedly contain variations, which allow us to consider them as different
versions of the same main process. In this experiment, we simulate the situation
where one version is in use, when a new version is suggested and AB-tested
5BPI Challenge 2015, including logs, reports, and process models:
https://www.win.tue.nl/bpi/doku.php?id=2015:challenge, Accessed 20-03-2017
Pre-print copy of the manuscript published by Springer
(available at link.springer.com)
identified by doi:10.1007/978-3-319-65000-5_7
10 S. Satyal, I. Weber, H. Paik, C. Di Ciccio, J. Mendling
Table 4: Analysis of versions A and B by cases
Metric Outcome Version A Version B Overall
Samples(N)
All 222 778 1000
Approved 72 275 347
Rejected 150 503 653
Median user satisfaction
All 3 3 3
Approved 3 5 5
Rejected 3 3 3
Median cost
All 795 1890 1890
Approved 2065 1990 1990
Rejected 795 1435 1435
Median duration
All 28.6 sec 17.4 sec 19.6 sec
Approved 35.5 sec 21.8 sec 22.4 sec
Rejected 24.5 sec 8.7 sec 8.9 sec
in competition with the previous one. Better performance here is equated to
shorter time to complete a process instance. Subsequently, the version that won
the first round competes against the next version, and so on, until all versions
have competed.
Based on the insights from [22], we filtered the events to retain only those
activity instances that belong to a so-called “phase”, namely constituting the
core part of the process. Using the Inductive Miner, we discovered five process
models P1, . . . , P 5from L1, . . . , L5, respectively. We mimicked the execution
of the processes by replaying the logs on the process versions. The instance
router decided to which alternative version to route each request to create a new
instance. In the following, we describe how the execution times were derived,
define the reward function, clarify how the competition was organized, and finally
report on the achieved results.
Execution time simulation. For fairness, in this experiment we replay only
the logs that did not stem from the original processes. Say, we are AB-testing Pi
vs. Pj; then we use the logs from Ltest “ tL1, . . . , L5uztLi, Ljuwith 1 ďi, j ď5.
However, we want to test how the event traces from Ltest behave on Piand Pj
in terms of timing. To this end, we assign an activity duration to the execution
of every replayed activity from the process version the activity was routed to.
Say this is Pi, and the current activity is a; then, we randomly sample a dura-
tion value from the durations of aamong all the traces of Lisharing the same
execution history (prefix) as the replayed trace.
For instance, consider the execution of activity phase concept draft decision
ready for process P1after the sub-trace rphase application received,phase ap-
plication receptive,phase advice known s. This activity has been assigned with
a random sample among the registered execution times of phase concept draft
decision ready-events following rphase application received,phase application
receptive,phase advice known sin the 593 traces of log L1sharing that prefix.
To that extent, we have folded the traces of every log into a dedicated poly-
tree auxiliary data structure, collapsing the traces that share the same prefix on
common paths. Every node keeps the activity name and the list of registered
Pre-print copy of the manuscript published by Springer
(available at link.springer.com)
identified by doi:10.1007/978-3-319-65000-5_7
AB-BPM: Performance-driven Instance Routing 11
execution times of the related events. In addition, event transitions in the log
are stored in a table structure along with the list of transition times. When the
traces cannot be followed in the poly-tree structure, we perform a lookup on the
table structure and derive execution times. If the current transition has not been
observed in Li, we discard the trace as non-conforming and disregard it in the
reward calculation.
The events in the logs signal the completion of an activity, and bear eight
timestamp fields. However, most of those attributes were missing or unreliable.
Therefore, we followed the approach of [22], and used solely the completion
time:timestamp attribute for each event. We computed the duration of every
activity as the difference between the timestamp of its completion and the pre-
ceding completion timestamp. We thus included in the activities’ duration esti-
mation both the execution time and the waiting time before starting.
Fig. 4: Reward strategy
Reward strategy. The filtered BPIC 2015
dataset contains numerous outliers: while the
median duration for processes are 39-46 days,
outliers can take up to 1158 days, i.e., 3
years and 63 days. To establish a reward func-
tion that penalizes very long process comple-
tion times, we adopted the following strat-
egy. Given the initial version Piof a pro-
cess, we collected all instance execution times
reported in its log Liand computed Kě1
quantiles q1, . . . , qK. We used these quantiles
to partition the space of possible execution
times into a set of K`2 intervals I“ tι0, ι1, . . . , ιK, ιK`1uwhere ι0“ r0, q0q,
ιK`1“ rqK,`8q, and ιk“ rqk´1, qkqfor 1 ďkďK. Those intervals split the
range of possible execution times for Pjas follows: ι0contains the values be-
low the minimum registered in Li,ιK`1accounts for any duration beyond the
maximum recorded in Li, and the Kintervals in-between are meant to classify
the performance of Pjwith respect to their quantile as per Li. This strategy is
illustrated as a step chart in Fig. 4 – every step in the chart represents a quantile.
The idea is to assign a reward of 1 when the duration Pjachieves is lower
than any registered execution time of Pi, and decrease it by a step of 2{Kas
long as the measured performance falls into the following intervals. In order to
counterbalance the disruptive effect of outliers which take extremely long, we
established a penalty value ρ(with ρ4inFig. 4) and defined that the reward
linearly decreases from ´1 to 1´ρqalong ιK. Finally, any execution time
beyond the last quantile is assigned a reward of 1´ρq.
Formally, let κIptq:R`Ñ r0, . . . , K `1sbe a mapping function associating a
process instance execution time tto the respective interval ιKPIby the index k,
e.g. κIptq “ 3 if tis in the range of ι3. The reward function rI:R`Ñ r´1´ρ, 1s
Pre-print copy of the manuscript published by Springer
(available at link.springer.com)
identified by doi:10.1007/978-3-319-65000-5_7
12 S. Satyal, I. Weber, H. Paik, C. Di Ciccio, J. Mendling
Algorithm 1: Strategy for the selection of the best performing version
among P1, P 2, P 3, P 4, P 5(.
1Ptest Ð P1, P 2, P 3, P 4, P 5(
2PiÐoriginal process version from Ptest
3Ptest ÐPtestz Pi(
4repeat
5PjÐalternative process version from Ptest
6Ptest ÐPtestz Pj(
7Ltest Ð tL1,...,L5uztLi, Lju
8PiÐbest version between Piand Pjas per the AB test over Ltest
9until Ptest ‰ H
10 return Pi
is defined over the set of intervals Ias follows:
rIptq “ $
&
%
1´κIptq
K¨2 if κIptq ă K
´1´ρ¨t´qK´1
qK´qK´1if κIptq “ K
´1´ρif κIptq ą K
with κIptq “
K`1
ÿ
k0
k¨χιkptq
where χιkis the characteristic function of interval ιk. The underlying idea was
to prefer the process demonstrating a shorter completion time to the slower ones
while accounting for the outliers. For our experiment, we set K20 and ρ4.
Competition: selecting the best version. To simulate the situation where
an organization gradually designs new versions of a process model, we run a
competition between the five provided process models. This competition is con-
ducted as a set of pair-wise comparisons between versions, following the schema
outlined in Algorithm 1. The idea is to initially consider an original version of
the process, Pi, and a new version, Pj. To determine if Pjachieves an actual
improvement over Piwhile limiting bias as discussed above, the execution of
the processes is simulated by replaying the traces in the logs from which Piand
Pjwere not derived. For instance, P1and P2are evaluated on the basis of the
traces in L3,L4, and L5. If, at the end of a competition round, Pjdemonstrated
an improvement over Pi, then Piis replaced with Pj. Otherwise, Piis main-
tained. At that stage, another process version is compared to Pi. The selection
procedure continues until all process versions have competed. We remark here
that the traces which could not be replayed on the process picked by the instance
router were discarded. The number of compliant traces still represents the vast
majority, because the ratio of conforming traces of all logs over models remained
around 0.9, and always above 0.7 as shown in Table 6. Also, the total number
of traces per log is shown in Table 5.
Analysis. Without loss of generality, we began the selection considering P1
as the process currently running on the production system, and progressively
entering P2,P3,P4, and P5into the competition as described above. We sped
up the execution time such that one day in the trace was equated to one second
in the experiments.
Pre-print copy of the manuscript published by Springer
(available at link.springer.com)
identified by doi:10.1007/978-3-319-65000-5_7
AB-BPM: Performance-driven Instance Routing 13
Table 5: Number of traces
Log Traces
L11199
L2830
L31409
L41051
L51155
Table 6: Ratio of conforming traces
Version L1L2L3L4L5
P11 0.928 0.949 0.974 0.928
P20.913 1 0.928 0.982 0.938
P30.901 0.812 1 0.975 0.886
P40.873 0.731 0.913 1 0.829
P50.897 0.929 0.944 0.979 1
Fig. 5: Request distribution over time
The sequence of tests was: (1) P1vs. P2,P2wins. (2) P2vs. P3,P3wins.
(3) P3vs. P4,P3wins. (4) P3vs. P5,P3wins. We can observe that P3was the
best-performing version. In all tests, the instance router chose the version with
lower mean and median execution time.
Fig. 5 shows the request distribution throughout the pair-wise tests. The
experimentation phase ends roughly after 1000 requests in all cases. We can
observe that occasionally the instance router decided to pick another version
some time during the post-experimentation phase. In some cases, the decision
made during the post-experimentation phase contradicted the decision during
the experimentation phase. In these scenarios, the instance router was able to
make the better decision only after all the delayed rewards were received.
In Table 7, we show the request distribution during the experimentation
phase, and the performance metrics calculated using execution times of processes
instantiated during this phase. Considering the median and mean times in this
table confirms that the instance router using the LTAvgR algorithm made the
right decision in all cases.
For an in-depth view of the reward estimates (the average reward observed
by LTAvgR) and execution times, we depict in Fig. 6 and Fig. 7 how their values
Pre-print copy of the manuscript published by Springer
(available at link.springer.com)
identified by doi:10.1007/978-3-319-65000-5_7
14 S. Satyal, I. Weber, H. Paik, C. Di Ciccio, J. Mendling
Table 7: Pair-wise performance comparison of versions after the AB Tests
Round 1 Round 2 Round 3 Round 4
Metric P1P2P2P3P3P4P3P5
No. of requests 559 440 423 575 263 729 735 261
Median duration 33.8 29.8 28.8 27 21 21.85 22.9 27.9
Mean duration 55.3 52.1 51.8 35.8 29.3 49.9 36.6 38.3
Fig. 6: Estimated rewards during the
experiment P1vs. P2
Fig. 7: PPI (duration) during the
warm-up phase, smoothed.
changed during the experiment P1vs. P2. The effect that fast completion leads
to positive rewards is clearly visible in Fig. 6: shortly after the start of the
experiment, the reward estimates for both versions jump to more than 0.6. After
some fluctuation, P1is preferred approximately from request 280 to request 811.
This is also visible in Fig. 5, where we can observe that the change in maximum
of the two reward estimates leads to change in the request distribution strategy.
Fig. 7 shows the PPIs observed by the instance router in order. Better PPIs,
which lead to better rewards, are received early. However, worse PPIs tend to
accumulate near the end of the warm-up phase. At request 811, the two estimated
rewards are very close to each other – see Fig. 6. At this point, P2collects actual
rewards from longer durations than P1– see Fig. 7. These longer durations result
in negative rewards, which cause the reward estimate of P2to fall below that of
P1. This development leads to the change in the decision of LTAvgR.
5 Discussion
The design of our routing algorithm, LTAvgR, was informed by practical ob-
servations of the limitations when applying existing algorithms in the process
execution context. As we have demonstrated, our approach addresses the key re-
quirements R1-R3. Our evaluation focused on the practical use of AB-BPM and
LTAvgR; theoretical analyses of the routing algorithm were out of the scope. The
in-depth analysis above showed how business-relevant PPI observations have a
direct influence on the routing decisions taken by LTAvgR.
We have used a multi-armed bandit algorithm with rewards derived from
a single PPI. In practice, however, multiple PPIs may need to be considered.
Pre-print copy of the manuscript published by Springer
(available at link.springer.com)
identified by doi:10.1007/978-3-319-65000-5_7
AB-BPM: Performance-driven Instance Routing 15
Furthermore, optimizing routing for one PPI can negatively affect other PPIs.
One key challenge in using multiple PPIs is that the reward delay for each PPI
can be different. Dealing with such scenarios may require improved collection
and reward update mechanisms, which we plan to explore in future work.
One limitation of our evaluation of AB-BPM so far is that they are based on
isolated environments with no real user interactions. Factors like effects from the
novelty of a process version were not considered. For example, in changing the
user interfaces and forms, we may observe that users behave differently when
exposed to a new version. As with the case study using real-world logs, we
expect to find some patterns unique to business processes when these factors are
accounted for. We believe that observations from real-world systems can guide
us towards designing a better instance routing algorithm, and identifying best
practices for performing AB tests on process versions.
6 Conclusion
Business process improvement ideas do not necessarily manifest in actual im-
provements. In this paper we proposed the AB-BPM approach which can rapidly
validate process improvement efforts, ensure fair comparison, and make process
level adjustments in production environments. Our approach uses an instance
router that dynamically selects process versions based on their historical per-
formance in terms of the chosen PPI. To this end, we proposed the LTAvgR
algorithm that can cater for the specifics of business process execution.
We evaluated our approach through synthetic and real-world data. We chose
the most effective routing algorithm based on a set of experiments, and evaluated
business process versions with synthetic data. Further, we evaluated real-world
process versions by performing pair-wise AB tests on them. The evaluation re-
sults showed that our instance router dynamically adjusted request distribution
to favour the better performing version.
In future work, we aim to integrate our framework with approaches to bal-
ance multiple PPIs. In addition, we plan to consider user interaction, and run
industrial case studies where we apply our instance router to actual production
systems.
Acknowledgements. The work of Claudio Di Ciccio has received funding from
the EU H2020 programme under the MSCA-RISE agreement 645751 (RISE BPM).
References
1. Agrawal, S., Goyal, N.: Thompson sampling for contextual bandits with linear
payoffs. In: Intl. Conf. Machine Learning, ICML (2013)
2. Alter, S.: Work system theory: overview of core concepts, extensions, and challenges
for the future. Journal of the Association for Information Systems p. 72 (2013)
3. Bass, L., Weber, I., Zhu, L.: DevOps: A Software Architect’s Perspective. Addison-
Wesley Professional (2015)
Pre-print copy of the manuscript published by Springer
(available at link.springer.com)
identified by doi:10.1007/978-3-319-65000-5_7
16 S. Satyal, I. Weber, H. Paik, C. Di Ciccio, J. Mendling
4. Burattin, A.: PLG2: multiperspective processes randomization and simulation for
online and offline settings. CoRR abs/1506.08415 (2015)
5. Burtini, G., Loeppky, J., Lawrence, R.: A survey of online experiment design with
the stochastic multi-armed bandit. CoRR abs/1510.00757 (2015)
6. Cabanillas, C., Di Ciccio, C., Mendling, J., Baumgrass, A.: Predictive task moni-
toring for business processes. In: BPM (2014)
7. Chapelle, O., Li, L.: An empirical evaluation of thompson sampling. In: Neural
Information Processing Systems (NIPS) (2011)
8. Crook, T., Frasca, B., Kohavi, R., Longbotham, R.: Seven pitfalls to avoid when
running controlled experiments on the web. In: ACM SIGKDD International Con-
ference on Knowledge Discovery and Data Mining. pp. 1105–1114 (2009)
9. Dumas, M., Rosa, M.L., Mendling, J., Reijers, H.A.: Fundamentals of Business
Process Management. Springer (2013)
10. Gregory, F.: Cause, effect, efficiency and soft systems models. J. Operational Re-
search Society pp. 333–344 (1993)
11. Hammer, M., Champy, J.: Reengineering the Corporation : A Manifesto for Busi-
ness Revolution. HarperCollins (1993)
12. Holland, C.W.: Breakthrough Business Results With MVT: A Fast, Cost-Free “Se-
cret Weapon” for Boosting Sales, Cutting Expenses, and Improving Any Business
Process. John Wiley & Sons (2005)
13. Jiang, W., Au, T., Tsui, K.L.: A statistical process control approach to business
activity monitoring. IIE Transactions 39(3), 235–249 (2007)
14. Kettinger, W.J., Teng, J.T., Guha, S.: Business process change: a study of method-
ologies, techniques, and tools. MIS quarterly pp. 55–80 (1997)
15. Kohavi, R., Longbotham, R., Sommerfield, D., Henne, R.M.: Controlled experi-
ments on the web: survey and practical guide. Data Min. Knowl. Discov. 18(1),
140–181 (2009)
16. Kohavi, R., Crook, T., Longbotham, R., Frasca, B., Henne, R., Ferres, J.L.,
Melamed, T.: Online experimentation at Microsoft. In: Workshop on Data Mining
Case Studies (2009)
17. Li, L., Chu, W., Langford, J., Schapire, R.E.: A contextual-bandit approach to
personalized news article recommendation. In: Intl. Conf. World Wide Web (2010)
18. Mann, H.B., Whitney, D.R.: On a test of whether one of two random variables
is stochastically larger than the other. The annals of mathematical statistics pp.
50–60 (1947)
19. ¯
Ono, T.: Toyota production system: beyond large-scale production. Productivity
press (1988)
20. Poelmans, S., Reijers, H.A., Recker, J.: Investigating the success of operational
business process management systems. Inf. Tech. Mgmt. 14(4), 295–314 (2013)
21. Sutton, R.S., Barto, A.G.: Introduction to Reinforcement Learning. MIT Press,
Cambridge, MA, USA, 1st edn. (1998)
22. Teinemaa, I., Leontjeva, A., Masing, K.O.: Bpic 2015: Diagnostics of building per-
mit application process in dutch municipalities. BPI Challenge Report 72 (2015)
23. Thompson, W.R.: On the likelihood that one unknown probability exceeds another
in view of the evidence of two samples. Biometrika 25(3/4), 285–294 (1933)
24. Weidlich, M., Ziekow, H., Gal, A., Mendling, J., Weske, M.: Optimizing event
pattern matching using business process models. IEEE Trans. Knowl. Data Eng.
26(11), 2759–2773 (2014)
25. Weidlich, M., Ziekow, H., Mendling, J., G¨unther, O., Weske, M., Desai, N.: Event-
based monitoring of process execution violations. In: BPM (2011)
Pre-print copy of the manuscript published by Springer
(available at link.springer.com)
identified by doi:10.1007/978-3-319-65000-5_7
This document is a pre-print copy of the manuscript
(Satyal et al. 2017)
published by Springer
(available at link.springer.com).
The final version of the paper is identified by doi:
10.1007/978-3-319-65000-5_7
References
Satyal, Suhrid, Ingo Weber, Hye-young Paik, Claudio Di Ciccio, and Jan
Mendling (2017). “AB-BPM: Performance-Driven Instance Routing for
Business Process Improvement”. In: BPM. Vol. 10445. Lecture Notes in
Computer Science. Springer, pp. 113–129. isbn: 978-3-319-64999-3. doi:
10.1007/978-3-319-65000-5_7.
BibTeX
@InProceedings{ Satyal.etal/BPM2017:ABBPM,
author = {Suhrid Satyal and Ingo Weber and Hye{-}young Paik and
Claudio {Di Ciccio} and Jan Mendling},
title = {{AB-BPM:} Performance-Driven Instance Routing for Business
Process Improvement},
booktitle = {BPM},
year = {2017},
pages = {113--129},
crossref = {BPM2017},
doi = {10.1007/978-3-319-65000-5_7}
}
@Proceedings{ BPM2017,
title = {Business Process Management - 15th International
Conference, {BPM} 2017, Barcelona, Spain, September 10-15,
2017, Proceedings},
year = {2017},
volume = {10445},
series = {Lecture Notes in Computer Science},
publisher = {Springer},
isbn = {978-3-319-64999-3},
doi = {10.1007/978-3-319-65000-5}
}
... Based on the performance of each version, the routing configuration can be dynamically adjusted to ultimately find the best performing version. This general idea has been introduced in AB-BPM [14]. However, the routing algorithm proposed in AB-BPM does not address scenarios where the process performance is measured through multiple Process Performance Indicators (PPIs) which may be available at different times. ...
... This anecdote shows that customer preferences are difficult to anticipate before deployment and that there is a need to carefully test improvement hypotheses in practice. Up until now, only the AB-BPM approach [14] supports the idea of using principles of AB testing to address these problems. As an early proposal, AB-BPM has a number of limitations regarding utilization of PPIs. ...
... We introduced the idea of AB testing for business process versions in AB-BPM [14], where we modeled this routing challenge as a contextual multi-armed bandit problem [4,12,2]. We proposed LtAvgR, which is based on LinUCB [12,5] -a well-known contextual multi-armed bandit algorithm. ...
Chapter
Full-text available
Business process improvement ideas can be validated through sequential experiment techniques like AB Testing. Such approaches have the inherent risk of exposing customers to an inferior process version, which is why the inferior version should be discarded as quickly as possible. In this paper, we propose a contextual multi-armed bandit algorithm that can observe the performance of process versions and dynamically adjust the routing policy so that the customers are directed to the version that can best serve them. Our algorithm learns the best routing policy in the presence of complications such as multiple process performance indicators, delays in indicator observation, incomplete or partial observations, and contextual factors. We also propose a pluggable architecture that supports such routing algorithms. We evaluate our approach with a case study. Furthermore, we demonstrate that our approach identifies the best routing policy given the process performance and that it scales horizontally.
... Based on the performance of each version, the routing configuration can be dynamically adjusted to ultimately find the best performing version. This general idea has been introduced in AB-BPM [14]. However, the routing algorithm proposed in AB-BPM does not address scenarios where the process performance is measured through multiple Process Performance Indicators (PPIs) which may be available at different times. ...
... This anecdote shows that customer preferences are difficult to anticipate before deployment and that there is a need to carefully test improvement hypotheses in practice. Up until now, only the AB-BPM approach [14] supports the idea of using principles of AB testing to address these problems. As an early proposal, AB-BPM has a number of limitations regarding utilization of PPIs. ...
... We introduced the idea of AB testing for business process versions in AB-BPM [14], where we modeled this routing challenge as a contextual multi-armed bandit problem [4,12,2]. We proposed LtAvgR, which is based on LinUCB [12,5] -a well-known contextual multi-armed bandit algorithm. ...
Conference Paper
Full-text available
Business process improvement ideas can be validated through sequential experiment techniques like AB Testing. Such approaches have the inherent risk of exposing customers to an inferior process version, which is why the inferior version should be discarded as quickly as possible. In this paper, we propose a contextual multi-armed bandit algorithm that can observe the performance of process versions and dynamically adjust the routing policy so that the customers are directed to the version that can best serve them. Our algorithm learns the best routing policy in the presence of complications such as multiple process performance indicators, delays in indicator observation, incomplete or partial observations, and contextual factors. We also propose a pluggable architecture that supports such routing algorithms. We evaluate our approach with a case study. Furthermore, we demonstrate that our approach identifies the best routing policy given the process performance and that it scales horizontally.
... Finally, there has been a recent attempt to use multi-armed bandit methods to replace A/B tests [23,24]. This techniques however, assume that one only has access to data from one of the experiments which differ from our full information scenario. ...
... It remains an open problem to provide a lower bound on the performance of our algorithm. [2,3], [4,7], [8,15], [16,23], [24,27] ...
Conference Paper
Full-text available
The rollout of new versions of a feature in modern applications is a manual multi-stage process, as the feature is released to ever larger groups of users, while its performance is carefully monitored. This kind of A/B testing is ubiquitous, but suboptimal, as the monitoring requires heavy human intervention, is not guaranteed to capture consistent, but short-term fluctuations in performance, and is inefficient, as better versions take a long time to reach the full population. In this work we formulate this question as that of expert learning, and give a new algorithm Follow-The-Best-Interval, FTBI, that works in dynamic, non-stationary environments. Our approach is practical, simple, and efficient, and has rigorous guarantees on its performance. Finally, we perform a thorough evaluation on synthetic and real world datasets and show that our approach outperforms current state-of-the-art methods.
Chapter
Business processes are complex and involve the execution of various steps using different resources that can be shared across various tasks. Processes require analysis and process owners need to constantly look for methods to improve process performance indicators. It is non-trivial to quantify the improvement of a proposed change, without implementing or conducting randomized controlled trials. In several cases, the cost and time for implementing and evaluating the benefits of these changes are high. To address this, we propose a principled framework using Structural Causal Models which formally codify existing cause-effect assumptions about the process, control confounding and answer “what if” questions with observational data. We formally define an end to end methodology which takes process execution logs and specified BPMN model as inputs for structural causal model discovery and for performing counterfactual reasoning. We show that exploiting the process specification for causal discovery automatically ensures the inclusion of subject matter expertise, and also provides an effective computational methodology. We illustrate the effectiveness of our approach by answering intervention and counterfactual questions on example process models.
Chapter
Full-text available
A fundamental assumption of improvement in Business Process Management (BPM) is that redesigns deliver refined and improved versions of business processes. These improvements can be validated online through sequential experiment techniques like AB Testing, as we have shown in earlier work. Such approaches have the inherent risk of exposing customers to an inferior process version during the early stages of the test. This risk can be managed by offline techniques like simulation. However, offline techniques do not validate the improvements because there is no user interaction with the new versions. In this paper, we propose a middle ground through shadow testing, which avoids the downsides of simulation and direct execution. In this approach, a new version is deployed and executed alongside the current version, but in such a way that the new version is hidden from the customers and process workers. Copies of user requests are partially simulated and partially executed by the new version as if it were running in the production. We present an architecture, algorithm, and implementation of the approach, which isolates new versions from production, facilitates fair comparison, and manages the overhead of running shadow tests. We demonstrate the efficacy of our technique by evaluating the executions of synthetic and realistic process redesigns.
Article
Full-text available
Adaptive and sequential experiment design is a well-studied area in numerous domains. We survey and synthesize the work of the online statistical learning paradigm referred to as multi-armed bandits integrating the existing research as a resource for a certain class of online experiments. We first explore the traditional stochastic model of a multi-armed bandit, then explore a taxonomic scheme of complications to that model, for each complication relating it to a specific requirement or consideration of the experiment design context. Finally, at the end of the paper, we present a table of known upper-bounds of regret for all studied algorithms providing both perspectives for future theoretical work and a decision-making tool for practitioners looking for theoretical guarantees.
Conference Paper
Full-text available
Information sources providing real-time status of physical objects have drastically increased in recent times. So far, research in business process monitoring has mainly focused on checking the completion of tasks. However, the availability of real-time information allows for a more detailed tracking of individual business tasks. This paper describes a framework for controlling the safe execution of tasks and signalling possible misbehaviours at runtime. It outlines a real use case on smart logistics and the preliminary results of its application.
Article
Full-text available
This paper presents a current, accessible, and overarching view of work system theory. WST is the core of an integrated body of theory that emerged from a long-term research project to develop a systems analysis and design method for business professionals called the work system method (WSM). After discussing WST's basic premises and its two central frameworks, this paper summarizes the relationship between WST and WSM. It shows how experience with early versions of WSM led to three extensions of WST that addressed limitations-in-use in one of the central frameworks in WST. After comparisons with related theories, this paper closes with an evaluation of progress to date, new directions for research related to WST, and implications for the IS discipline. The two appendices summarize the long term research from which WST emerged and use a positioning map to show how WST is related to other topics in the IS discipline.
Article
Full-text available
Business process management systems (BPMS) belong to a class of enterprise information systems that are characterized by the dependence on explicitly modeled process logic. Through the process logic, it is relatively easy to manage explicitly the routing and allocation of work items along a business process through the system. Inspired by the DeLone and McLean framework, we theorize that these process-aware system features are important attributes of system quality, which in turn will elevate key user evaluations such as perceived usefulness, and usage satisfaction. We examine this theoretical model using data collected from four different, mostly mature BPM system projects. Our findings validate the importance of input quality as well as allocation and routing attributes as antecedents of system quality, which, in turn, determines both usefulness and satisfaction with the system. We further demonstrate how service quality and workflow dependency are significant precursors to perceived usefulness. Our results suggest the appropriateness of a multidimensional conception of system quality for future research, and provide important design-oriented advice for the design and configuration of BPMSs.
Article
Process mining represents as an important field in BPM and data mining research. Recently, it has gained importance also for practitioners: more and more companies are creating business process intelligence solutions. The evaluation of process mining algorithms requires, as any other data mining task, the availability of large amount of real-world data. Despite the increasing availability of such datasets, they are affected by many limitations, in primis the absence of a "gold standard" (i.e., the reference model). This paper extends an approach, already available in the literature, for the generation of random processes. Novelties have been introduces throughout the work and, in particular, they involve the complete support for multiperspective models and logs (i.e., the control-flow perspective is enriched with time and data information) and for online settings (i.e., generation of multiperspective event streams and concept drifts). The proposed new framework is able to almost entirely cover the spectrum of possible scenarios that can be observed in the real-world. The proposed approach is implemented as a publicly available Java application, with a set of APIs for the programmatic execution of experiments.
Article
A growing number of enterprises use complex event processing for monitoring and controlling their operations, while business process models are used to document working procedures. In this work, we propose a comprehensive method for complex event processing optimization using business process models. Our proposed method is based on the extraction of behavioral constraints that are used, in turn, to rewrite patterns for event detection, and select and transform execution plans. We offer a set of rewriting rules that is shown to be complete with respect to the all, seq, and any patterns. The effectiveness of our method is demonstrated in an experimental evaluation with a large number of processes from an insurance company. We illustrate that the proposed optimization leads to significant savings in query processing. By integrating the optimization in state-of-the-art systems for event pattern matching, we demonstrate that these savings materialize in different technical infrastructures and can be combined with existing optimization techniques.
Article
Thompson sampling is one of oldest heuristic to address the exploration / ex-ploitation trade-off, but it is surprisingly unpopular in the literature. We present here some empirical results using Thompson sampling on simulated and real data, and show that it is highly competitive. And since this heuristic is very easy to implement, we argue that it should be part of the standard baselines to compare against.
Article
Statistical Process Control (SPC) techniques have been successfully used in manufacturing industries to trigger and identify the root cause of variations so as to promote quality improvement. This paper develops a SPC framework to identify important changes deserved in business activity monitoring. To model and track thousands of diversified customer behaviors, the proposed SPC system consists of efficient and robust profiling methods to accommodate different behavior patterns including business changes, structural breakdowns, and unnecessary errors. Several customer profiling techniques are discussed and the activity monitoring performance based on the profiling algorithms is compared in a simulation example and a customer churn detection example in a telecommunications setting. The enhanced system will allow business managers and engineers to establish successful customer loyalty programs for churn prevention and fraud detection.
Article
The logical connectives in the conceptual models of Soft Systems Methodology are limited to relations of 'necessity'. This is not enough to achieve a correspondence with states of affairs in the physical world. In order to attain this correspondence connectives representing 'sufficiency' must be included. When this is done a logical account of efficiency is possible. This defines efficiency as the arbiter between two or more sufficient but unnecessary conditions of a desired effect.