Available via license: CC BY-NC 4.0
Content may be subject to copyright.
Case-Intelligence Recommendation on Massive
Contents Processing through Dynamic
Computing
Rui Li
Department of Information Engineering
Anhui Communications Technical College
Hefei, China
Liruilary@gmail.com
Jianyang Li, Benkun Zhu
School of Computer and Information
Hefei University of Technology
Hefei, China
lijianyang@sina.com
Abstract—How to suggest a valid recommend within a
reasonable time is the greatest technical challenge for the
recommendation system, for which tremendous user cases
with high dimension are generated while it runs in real time,
and these massive data are too difficult to compute directly.
This paper proposes a case -intelligence system framework
along with a feature -based multi -layer feed -forward neural
networks (MFNN) to succeed case- retrieval based on
dynamic computing, which constructs the neural networks
dependence on the real input vectors instead of the fixed and
dull networks structure presupposed, and can apply many
kinds of knowledge granularity from various levels
effectively to help users for information retrieval and case
adaptation. Our subsequent experimental results indicate
that it is capable of handling the massive personalized data,
and our covering algorithm can decrease the complexity of
MFNN algorithm for dynamic computing, which performs
adaptable knowledge granularity to enhance the system's
efficiency of reasoning.
Keywords- case-intelligence recommender; dynamic
computing; covering algorithm; MFNN; system efficiency
I. INTRODUCTION
The acquisition of personalized need is the key to
effective recommender, which can capture user's
information demand exactly, and promote the wide
application of E-commerce. Many intelligent tools have
been used to help users search, locate and manage web
documents, such as data mining and other artificial
intelligence technology used to collect data, obtain their
behaviors in e-commerce, and generate interests in the
products for consumer’s purchase [1]. Thus, more and
more recommendation systems have been developed to fit
for e-commerce use [2], which could be distinguished into
three different types: rule-based filtering, content based
filtering, and collaborative filtering. Personalized
information acquirement is an important research center
naturally, for they realize that efficient access and quality
of service are helpful to attract more visitors [3].
But the statistics report from ACM indicates that the
current recommender can not meet the large-scale e-
commerce applications, and it has poor real-time problem
accompany with the problem of weak quality in accuracy.
As well known, the personalized information- the real user
behavior data from websites, can accumulate up to
millions or even billions for the recommendation system
running, the processing of massive user data is the greatest
challenge, for which involves system performance deeply
[4]. Because of the recommended system is a data priority,
the more accumulation of data, and the higher accuracy the
recommender can explore [5]. How to suggest a valid
recommend in a reasonable time people need from the
mass merchandise has become increasingly difficult,
which is exactly the same with case intelligence as we
have described in the paper [6].
Normally, recommendation system uses low level
analogy reasoning, which is an important cognitive model
of human sense [7]. The “low level” means simple analogy,
which is not inferring in different domain data. For the
incomplete knowledge implicit in the reasoning, the
conclusions of analogy may be effective or invalid, which
must require an objective confirmation or readjust until
new knowledge or contradiction comes. Generally, CBR
system implements four processes well known as the 4R -
Retrieve Reuse, Revise, and Retain, to solve new problems
[8]. The former cases can also be used to evaluate the new
issues and new programs of problem-solving [9], and
prevent the potential errors in the future. Cases can be
reused by similarity computing as case knowledge space
conversion, which is the glorious with exciting highlight in
the construction of CBR intelligent system, and the
characteristic advantage distinguishes CBR systems from
RBR systems thoroughly [10].
System flexibility depends on case knowledge space
conversion through case- adaptation method, whose
process is manipulate the adjustment of space projection
based on former knowledge; So that the system outputs a
set of most similar cases to guide users’ corresponding
inputs, from which case- adaptation can get great benefits
[11]. This paper focus on such problems by using MFNN
to acquire flexible knowledge for our synthesis reasoning,
which mainly comes from Granular Computing and Case-
Based Reasoning, can combine various reasoning
principles and integrate many methods, and enhance the
system's efficiency of reasoning by means of dynamic
granularity.
II. CASE SELECTION MODELING
The acquisition of user’s data is the key to meet the
individual needs and infer the learning ability of the
recommender. How to obtain personalized demands,
expression knowledge to support user from information
retrieving and adapting, is the most important task to
intelligent recommendation system. The paper [12] has
International Conference on Mechatronics, Electronic, Industrial and Control Engineering (MEIC 2014)
© 2014. The authors - Published by Atlantis Press
1060
described that case-intelligence recommendation system
can be used for acquiring effective personalized
knowledge, besides several other adaptive interfaces
attempt to collect user information unobtrusively.
A. Case retrieval
Case-Based Reasoning as a cognitive model suggests
people learn the best from former problem-solving cases as
they solve new problems, which is the simulator of human
analogy learning. Case retrieval is the key process of the
CBR system, and plays an important role in Machine
Learning community. Case is the integrated representation
of human sense, logics and creativity, great achievement
has been acquired for CBR in the field of knowledge lack,
and case intelligent decision techniques is built from CBR
can overcome some defects, such as poor flexibility, and
provides decision support [13].
Artificial Neural Networks has the natural relationship
with CBR, and several successful theories have been put
forward to integrate ANN into the CBR system. Similarity
assessment plays a key role in lazy learning methods, but
the traditional k-nearest neighbor (kNN), which are
applicable to any representation for cases gathered,
measures similarity between cases are time-consuming.
Specifically, the similarity measurement is empirically
evaluated on relational data sets of different expressiveness
[14]. The case library in the CBR system can be viewed as
a CSP, therefore CS-ANN model, such as Schema model,
Hopfield model, Boltzmann and Harmony theory can be
employed to construct the case library. Theoretically, in
the symbolic description model-based CBR system, rules
can be elicited by ANN method; and in the quantitative
description model-based CBR system, due to the system’s
flexibility, many mathematical approaches and
optimization techniques can be employed in the definition
and analysis of similarity measurement and case adaptation
criteria, thereby more and more ANN applications can be
prevailed in the CBR system.
There have been a lot of very wide and profound
researches on this topic, including data integration, query
processing, and fine-granularity data sharing. Currently the
main way for case-matching is the k-nearest neighbor
algorithm, but it can not reflect the relationship between
the cases and as well as their attributes, neither can it
shows the preference of the customers. The widely used
BP Network can be used to create a CBR retrieval model
and its most outstanding characteristic is that the retrieving
speed and the size of the case library are in a non-linear
relationship. But some insurmountable weaknesses
remained in these application systems, for example, the
weak performance in interpretation and large-scale case
library that due to the high complexity of ANN algorithm
makes the systems far too complex and hard to be
integrated. Especially for the large-scale case library, the
retrieving time is unacceptable. Facing these problems, we
should employ the MFNN dynamic computing instead.
After investigating the behavior of MFNN together
with many kinds of existent algorithms for case retrieval
based on MFNN, besides BP, simulated annealing
algorithm and their ameliorated algorithms, we found that
weaknesses such as having lower speed and local extreme
value, are inherent in those algorithms, and cannot be
conquered thoroughly. Considering that more and more
interests are focusing on data intensive computing and data
cloud computing in industry and academia, those methods
can not be used in large- scaled case library especially for
dynamic computing as intelligent case retrieval techniques.
In this paper, we suggest to use MFNN and employ
Covering algorithm [15], which is easily understandable
and constructed, to effectively decrease the complexity of
ANN algorithm, to manipulate the massive personalized
users’ data in real time.
B. Personalized features
In order to achieve personalization services, we must
trace down user’s behavior to study his interest, which can
be collected from three sources: Server, Client, and Proxy,
to exploit potential information or patterns useful. There
are three types log files to record users’ actions: Access
logs, Refer logs and Agent logs, also may be Cookie logs.
Besides, there are query information, register information,
and website structure. These data can be divided into the
following categories:
Content data: the real data which user having read
and used, mainly constituted by text and image.
Structure data: describing how to construct
website and organize the webpage. The page can
be constructed by HTML, XML representation as
the tree structure, and HTML label set as the root
for the tree, whose structure can be connected by
hyperlinks between the different pages; and web
structure data mining refers to a method of mining
the structure between different web pages user
browsing.
Usage data: describing web usage pattern, such as
IP address, URL, webpage citation, access time
and date, which indicates each user’s behavior
model, and a typical use of data originates from
server log. The Usage Data Collector (UDC) is a
framework for collecting usage data information,
which gathers information about the kinds of
things that the user is doing (i.e. activating views,
editors, etc.).
User Profile: relative statistics data from web user,
including user registration and personal
information, for example, user name, education
background, career, position, age, income,
personal hobby, etc. Due to the dynamic and
complex nature of web users, automatically
acquiring user profiles is very challenging.
Recommender is running in much complex
environment, each data is regarded as the foundation for
knowledge representation, but may be represented in semi-
structured or unstructured model, or even in natural
language texts. Although the research of personalized
interest has become widespread only in recent years,
several adaptive interfaces have been developed to
describe personalization by observing a user's browsing
behavior for promoting recommendation performance.
Web data mining has been widely used for sharing and
exchanging of data and resources among numerous
computer nodes, and personalized data objects could be
identified with high-dimensional feature vectors. Though
an investment table feed back by user may be acquired,
user behavior extraction is compulsory to estimating the
user interest degree. There is a growing trend among
1061
companies,
organizations and
individuals to gather
information through web
data mining to utilize
that information in their
best interest [16]. TM is
the topic-keywords
matrix of user’s weight for the expression of user interest
degree in our paper, where n represents the number of user
interested topics, and w is the weight. A new approach of
web DM technologies to users’ data can generate an
analysis of customers’ behavior, by synthesizing key
abstract information that will facilitate and improve the
customization of services.
C. System construction
The personalized recommendation system that we
proposed based on case intelligence mainly construct by
three parts: input module, recommendation methods and
output module, as shown in Figure 1, which the dynamic
computing module with MFNN algorithm as we have
described, is added in as retrieve process. Our dynamic
computing method will be validated in the sequent chapter,
which will be described in detail and be evaluated directly
with huge personalized users’ data.
Figure 1. the system frame
Personalized recommendation system involves a serial
processes of gathering and storing information about site
visitors, managing the content assets, analyzing current
and past user interactive behavior, and delivering the right
content to each visitor based on its analysis, the details of
whose components is not described in this paper, which
can be seen in paper [6]. We care only about the dynamic
computing method prepared for case matching, revealing
and exploiting the most similar users’ groups, where the
implementation process of personalized recommendation
for the same with common recommendation system is also
not mentioned.
III. DYNAMIC COMPUTING FOR RECOMMENDER
Such personalized data are difficult for computing, for
they are massive with high dimensional and increasing in
every moment drilled from websites. While the
accumulation of the real user behavior data up to millions
or even billions in real time, the greatest technical
challenge most organizations face is how to suggest a valid
recommend within a reasonable time from the dynamic
data-ocean.
A. Dynamic computing algorithm
MFNN consists of an input layer, one or more hidden
layers, an output layer, where layers are in order of priority,
and the i-layer neurons receive signals only from the i-1
layer neurons without feedback between each layer. It has
proved that 3-layer networks can realize any given
function for approximate accuracy, and can be used to
solve the nonlinear classification. This paper suggests
Multi-Layer covering algorithm for dynamic computing,
which is a constructive method for ANN, and the
foundation of the geometrical representation McCulloch-
Pitts neural model.
Our networks construct its layer-structure by means of
input training data for itself. Firstly, Covering Algorithm
assumes that each input vectors
x
of an n-dimension can
be projected on a bounded set of a certain hyper-sphere
n
S
of a (n+l)-dimensional space (define the sphere radii is
R), it is no doubt that the transformation must be achieved
to the aim through widening the vector’s dimension. The
dynamic algorithm is described as the following steps:
(1)Search for the maximum sample ||r|| from the
learning samples X, then project all the points in X to the
sphere, which centers at the base point, radius R(R> = r);
(2)Assume i=1, t=0; // i represents the i-th class and
regarded as the sample covering center for covering; t is
the number of covering domain;
(3)Then let m=mod(i, N):
If
m
X
=, goto(11),
Else t++, randomly select a point
i
a
from
m
X
;
(4)Calculate
i
d
=
},{max
xai
to make the cover
C(
i
a
), which centers from
i
a
, θ as the threshold;
(5)Calculate the barycenter of all the points in C(
i
a
),
project it to the sphere and get projective point
'i
a
, then
calculate its threshold
'
and partition the sphere domain
C(
'i
a
).
(6)If the total points which C(
'i
a
)covers, is more than
what C(
i
a
)covers,
Then let
'i
a
i
a
，
'
， return(5),
Else restore the original parameters;
(7)Calculate
),(|min x
adxB
m
Xj
,
11 12 1
21 22 2
12
...
...
... ... ... ...
...
m
m
n n nm
w w w
w w w
TM
w w w
Rating Base
Adapted Knowledge
Base
Case Base
User Data
Base
Similar cases
Proposed
solution
Confirmed solution
Retrieve
Reuse
Revise
Problem
Various Artificial
Intelligent Modules
Retain
Input Module
Output
Module
Dynamic
Computing
1062
Where d(a, x)represents the distance between a and x,
let the translation of point a '= a;
(8)If |B|>n, (|B| represents the card(B)), goto(10)
Else find the pedal b from a to P(B), let P(K)= P(B),
for each xX /(P(K)∪Xm)，
calculate d(x)：d(x)=<a,c-x>/<b,c-x>
Where c is the random point P(K).
If there exists x: <a, cx> = 0,
then let ck + 1 = x, a '= a, P(K + 1)= P(K)∪ {ck + 1}
else let d =d(x*)=
m
Xx
min
{d(x)},
assume a′=R(a-db)/|a-db|，
where R is the radius of the spherical Sn+1
project vector(a-db)to Sn+1, take ck + 1 =. x *.
(9)P(k+1)= P(k)∪ {ck+1}(P(k)∪ {ck +1}
if k+1 > n, then a' is the result, and goto(11)
else project a' to P(k + 1), b=bk+1 ，a= a',
k++ ,return(8)
(at the beginning, let k = |B|).
(10)Count for the corresponding spherical domain
C(
'i
a
)
If it is more than what C(
i
a
)covers,
let
'i
a
i
a
，
'
， return(5)
Else restore the original parameters, and get a covering
domain C(t)of
m
X
;marked with Cp(t) ，
m
X
/ Cp(t)
m
X
(11)Count for nonempty set among X1, X2, ..., XN,
If the number is greater than 1; i++; return(3);
By this way of dimension expansion and space
projection, the domain covering for the similar users in the
user library will be well achieved, which can be used as the
input of the MFNN for case matching in dynamic
computing.
B. Experiments with outlook
Two experiments are designed to validate our system
algorithm, and the experimental data of “forest cover type”
is downloaded from UCI repository, whose main
information describes as follows: Number of instances
(observations) 581012, Number of Attribute: 54 (12
measures, 10 quantitative variables, 4 binary wilderness
areas and 40 binary soil type variables); Number of Class:
7; Missing Attribute: none. Each record represents the user
personalized data collected from the websites, which is
regarded as a user behavior vector with 54 Attributes, and
user data library accumulates to 581,012 users’ sessions in
real time. Then, the normal Macro-averaging is used to
calculate all classes’ means
Fscore
:
2* *recall precision
F score recall precision
We can find the data are spare matrix with high
dimensional and huge records. The actual forest cover type
for a given observation (30 x 30 meter cell) is determined
from US Forest Service (USFS) Region 2 Resource
Information System (RIS) data; each record is regarded as
a user case in our experiments, to decrease the interference
in the pretreatment of the real world. The simulation starts
from 10,000 to 100,000 user cases, and the user cases are
randomly selected.
The first is designed to confirm that our MFNN and its
algorithm are reliable for its excellent precision and
outlook speed, as Table 1 shows. We can find that each
covering domain is a most similar group, and can be
represented as a most similar user group in the same
interest degree, which can be recommended for the new
user with the same interest. Besides, user data
accumulating is gradually adding up, which can be
calculating in the background, and we can use the results
of the previous partition without recalculating the groups.
TABLE I. SYSTEM PERFORMANCE
The subsequent experiment is to validate it for large-
scale data in dynamic computing. Original user records are
increasing with a serial of new users adding in, where
original records represent previous personalized resource
that the recommendation system have drilled from
websites formerly and saved as the greatest asset. While
new users’ data is adding in the system, this is just like
what the new personalized data is acquired and stored in
case library in real time. So the second experiment is
divided into two items to simulate dynamic users’ action,
the one is set with the fixed 2,000 new users adding in, the
other with the fixed 10,000 with comparison.
TABLE II. SYSTEM COST IN REAL TIME
As Table 2 indicates, it costs only a little system
resource to recommend the most similar users’ case in
dynamic computing, and states clearly that the cost for
system data recalculating is valuable and acceptable.
Though the real commendation system cannot recalculate
its personalized data library unless system collapsing, the
costs in our experiments for new users’ data feed in the
library are under recalculating to meet the general testing
of Machine Learning algorithm demands.
Records Domains F-score(%) T-partion(s) Time(ms)
10,000 2468 79.1 14.207 14.81
15,000 3607 81.7 33.225 34.339
20,000 4360 82.9 49.209 50.391
30,000 5646 81.3 68.316 69.577
40,000 6391 82.4 86.05 87.349
50,000 6843 81.6 103.06 104.4
100,000 7651 83.2 272.31 273.75
Origin ⊿Records ⊿Domains ⊿T-partion(s) ⊿Time(ms)
2,000 2,000 414 61.2 1.5895
4,000 2,000 434 69.1 3.2935
6,000 2,000 545 74.2 3.8923
8,000 2,000 575 79.1 4.997
10,000 10,000 1,139 81.7 19.529
15,000 10,000 753 82.9 16.052
20,000 10,000 1,286 81.3 19.186
30,000 10,000 745 82.4 17.772
40,000 10,000 452 81.6 17.051
1063
IV. CONCLUSION
Personalized data which are explored from websites,
are the greatest asset for recommendation system, but they
are massive along with high dimension, and hardly dealt
with in real time, especially in dynamic environment.
Addressing such problem of efficiency, the paper suggests
an idea of intelligent information retrieval processing,
whose basic task is to construct a suitable granularity to
decrease the system complexity for real time computing;
and establishes a user model for personalized
recommender based on our dynamic computing algorithm,
which has such advantages like clear system structure,
feasible component combination, and can effectively help
users for information retrieval and case adaptation.
ACKNOWLEDGMENT
This research is supported by the Natural Science
Project of Anhui Province under grants KJ2014A050.
REFERENCES
[1] Wei Chu，Seung-Taek. Park, “Personalized Recommendation on
Dynamic Content Using Predictive Bilinear Models”, WWW2009,
pp691-700
[2] Khalid Al-Kofahi, Peter Jackson etc, “A Document
Recommendation System Blending Retrieval and Categorization
Technologies”, AAAI Workshop 2007, pp 9-18
[3] Zurina Saaya, Markus Schaal, Maurice Coyle, Peter Briggs, and
Barry Smyth. “Exploiting Extended Search Sessions for
Recommending Search Experiences in the Social Web”. ICCBR
2012. LNAI, (7466), pp. 369-383
[4] Zurina Saaya, Markus Schaal, Maurice Coyle, Peter Briggs, and
Barry Smyth, “Exploiting Extended Search Sessions for
Recommending Search Experiences in the Social Web”, ICCBR
2012. LNAI, (7466), pp369-383
[5] Ruihai Dong, Markus Schaal, Michael P. O’Mahony, Kevin
McCarthy,and Barry Smyth, “Opinionated Product
Recommendation”, ICCBR2013, LNCS,(7969), pp44-58
[6] Jianyang Li, Xiaoping Liu. “A Case-intelligence Recommendation
System on Massive Contents Processing through RS and RBF”.
ICMTMA 2013. pp1-4
[7] Amira Abdel-Aziz, Weiwei Cheng, Marc Strickert, and Eyke
Hüllermeier, “Preference-Based CBR: A Search-Based Problem
Solving Framework”, ICCBR 2013, pp1-14
[8] Odd Erik Gundersen, “Toward Measuring the Similarity of
Complex Event Sequences in Real-Time”, ICCBR 2012. LNAI,
(7466), pp.107-121
[9] Debarun Kar, Sutanu Chakraborti, and Balaraman Ravindran,
“Feature Weighting and Confidence Based Prediction for Case
Based Reasoning Systems” , LNAI, (7466), pp. 211-225
[10] Jianyang Li,Xiaoping Liu, Rui Li. “Application of Improved
MFNN on Dynamic Computing for Case-Intelligence
Recommendation System”. IMSNA2012, pp407-410
[11] Bach, K., Althoff, K.-D., Newo, R., Stahl, A. “ A Case-Based
Reasoning Approach for Providing Machine Diagnosis from
Service Reports”. ICCBR 2011. LNCS, (6880), pp. 363-377
[12] Jianyang Li, Xiaoping Liu. “Personalized Recommendation
System on Massive Content Processing Using Improved MFNN”.
Springer's LNCS, 7529 (2012), pp183-190
[13] Zhiwei Ni，Jianyang Li，Fenggang Li，Shanlin Yang. “Survey
of Case Decision Techniques and Case Decision Support System”.
Chinese Computer science,2009,36(11), pp18-24
[14] David McSherry, “Conversational Case-Based Reasoning in
Medical Decision Making”, Artificial Intelligence in Medicine,
(52), 59–66 (2011)
[15] ZHANG Ling. “The relationship between Kernel Functions Based
SVM and Three-layer Feedforward Neural Networks”. Chinese J.
Computer, 25(7): 696-700, 2002.
[16] Catherine Havasi, Jason Alonso, Robert Speer, “Reducing the
Dimensionality of Data Streams using Common Sense”,
WWW2010, pp 14-19
1064
