Abstract—With the fast development of information
technology, retailers are suffering from the excess of information.
Too much information can be a problem. However, more
information creates more opportunity. In retailing, information is
the key issue to maximizing revenue. It is now hard to make timely
or effective decisions and to the right content to the right place, at
the right time and in the right form. This paper is about managing
the information so that the user can gain more clear insight. It is
about integrating and inventing methods and techniques. The
Semantic Web will provide a foundation for such a solution.
However, semantics only provide a way of mapping the content of
a web to user defined annotations. Not many companies have fully
utilized the power of Internet retailing due to the various technical
obstacles have yet to be overcome. The existing research in
e-retailing focuses only on the traditional retailing including
direct and indirect retailing approaches. This paper suggests that
applying association mining techniques can further improve the
dealing of information overload in a web oriented retailing
environment.
Index terms—Semantic web, online retailing, data mining,
formal concept, Protégé, triple store, Sparql.
I. I
NTRODUCTION
nformation is all around us, easy to collect, store and access.
It consists of the useful data that is needed to solve our
problems. But as information is more and more overloaded,
managers, researchers or retailers have to spend more time to
process the information before making their decisions, due to
the fact that the stored information is unstructured.
One way of solving this problem is to turn the important and
useful information into knowledge and filter away the less
important information. This requires understanding of when to
use information, how to find it, and how to present it to the
target customer. It is imperatives that enterprises will need to
exploit the knowledge and information available on the WWW
as far as possible, so its static nature will tend to increase the
information overload. This paper suggests using Semantic Web
to solve the information overload problem. The technological
basis of the Semantic Web provides a unified framework within
which many approaches to the problem of information overload
can be integrated.
Manuscript received May 10, 2008. Manuscript accepted for publication
June 20, 2008.
Yi Wang is with Nottingham Business School, Park Row, floor 2,
Nottingham Trent University Burton Street, Nottingham, NG1 4BU, U.K.
The major components in Semantic Web are shared
conceptualizations and terminologies to describe customers,
operation, content of web page, etc. These conceptualizations
and terminologies are called ontologies. They may refer to
agreed ways of describing customer’s preference and demand,
operational capacity and constraints, retailing brand.
Ontologies can be used as meaningful enrichment for the
content. They provide common base framework within which
information can be properly shared, modeled and filtered. The
next part of paper reviews some existing literature regarding
online retailing and information overload. The third part of the
paper explains semantic web as a mean to solve information
overload. The fourth part gives some technical background for
the usage of semantic web in retailing. The fifth part suggests
some improvement based on existing technologies. The sixth
part will discuss in detail about dynamic causal mining as a
specific data mining tool. The seventh part will illustrate a
practical problem. The final part concludes the whole paper.
II. L
ITERATURE REVIEW
On-line retailing offers more choices and flexibility
[Lamoureux, 97] and, at the same time, eliminates huge
inventories, storage costs, utilities, space rental, etc. [Avery,
97]. Companies can design and personalize advertising for each
customer [Peterson, et al., 1997]. The Internet can provide
timely information to customers because of its ability for
instant communication [Lane, 1996]. This means more
interaction [Rosenspan, 2001] and quicker responses [Isaac,
1998]. The information can be used to assist new product
development and introduction [Higgins, 2001]. The
communication also helps with identifying prospects [Ebling,
2001], sales and relationship building [Ginovsky, 2001], and
deepening customer loyalty [Kranzley, 2001]. Perlow [1999]
describes a software company characterized by an
environment. It also allows for easy follow-up on customers’
needs [Marks, 1998]. Retailing activity occurs through three
types of channels: communication, transaction, and distribution
channels [Peterson, 1997].
Studies of the semantics web were initiated by Tim
Berners-Lee, the creator of the World Wide Web [Berners-Lee,
01]. The Web is referred to as the “semantic Web”, where
information will be retrieve from intelligent network services
such as information brokers and search agents [Decker &
Melnik, 00, Decker et al., 00].
The World Wide Web has evolved into a dynamic,
distributed, heterogeneous, complex network, which is hard to
Yi Wang
Applying Dynamic Causal Mining
in Retailing
I
Yi Wang
control [Albert et al., 99, Huberman & Adamic, 99]. It is
important to have consistent understanding and interpretation
[Helbing et al., 00,]. in the World Wide Web [Huberman &
Adamic, 99, Huberman et al., 97, Barabasi & Albert, 99].
When more information arrives than individuals can process,
an information overload occurs [Simon, 1971]. Much research
has done in dealing with information overload. Abiteboul et al.
[Abiteboul, 00] systematically investigated the data on the Web
and the features of semistructured data. Zhong studied text
mining on the Web including automatic construction of
ontology and filtering system [Zhong, 00; Zhong et al., 00b].
Liu et al. worked on e-commerce agents [Liu & Ye, 01] and
KDDA (Knowledge Discovery and Data Mining Agents) [Liu
& Zhong, 99, Liu et al, 01] to minimize the information
overload.
III. A
GENERAL FRAMEWORK OF SEMANTIC WEB
Figure 1 shows the simplified representation of a semantic
network where enter prise on one side provides information and
the user on the other hands give the queries.
The semantic network interface has three goals:
1. To provide ontologies for interoperations. In many cases,
the customer and the enterprise does not speak the same
language. They have different preference, gain different
knowledge and get different information from the product or
services. In order to have a more successful relationship, there
is need for some ontology which bridges these gaps.
2. To unify information from different document format.
The enterprises provides the information online in different
format, some as word documents, some as excel documents,
some as media files., etc. The goal here is to integrate all
relevant based on the given ontology. And map all the
information to a user friendly representation.
3. To Store the relevant information and update the
ontology. The interface should be able to retrieve and represent
the information based on user’s queries. And the interface
should be able to update the ontologies for improvement, thus
rather storing a large amount of information, the relevant
ontologies or relations are stored.
IV. O
NTOLOGY DEVELOPMENT AND STORAGE
An ontology for retailing defines a common vocabulary for
any participants in retailing, including customer, manager,
retailer, etc. who need to share information in a domain. It
includes machine-interpretable definitions of basic concepts in
the domain and relations among them. The major goals for
ontology development (Natalya and McGuinness, 07) are:
1. To share common understanding of information among user
(Musen 1992; Gruber 1993).
2. To enable reuse of domain knowledge was one of the driving
forces behind recent surge in ontology research.
3. To change domain assumptions if information about the
domain changes.
4. To separate the domain information from the operational
information
5. To reuse the existing ontologies and extending them
(McGuinness et al. 2000).
The most common tools for developing ontology are
Protégé-2000 (Protege, 2000), Ontolingua (Ontolingua, 1997),
and Chimaera (Chimaera, 2000) as ontology-editing
environments.
Figure 2 shows a sample of the protégé interface for wine
ontology. It is typically based on graphical class hierarchical
development (Uschold and Gruninger, 96). An ontology can be
seen as a triplets (Subject, relation, object). An typical example
is “Wine”, which is subject, “has brand name”, which is a
relation, “Chinati” which is an object. Instead of normal Sql
technology, the ontology can be store in a triple store, which is
a specific type of data storage. It is designed to store and
retrieve identities that are constructed from triplex collections
of strings (sequences of letters) and can be queried with
Sparql. Sparql query consists of three parts. The pattern
matching part, which focus on matching patterns of graphs, like
optional parts, union of patterns, nesting, The solution
modifiers part , which allows to modify values applying
classical operators like projection, distinct, order, limit, and
offset. And the output part which consists of different types:
yes/no solution based on descriptions of resources.
Applying Dynamic Causal Mining in Retailing
V. MINING TECHNIQUES
The ontologies developed for semantic web are based on
user experience. This requires the need for the developers to
have an increased understanding of the complex issues
involved in the ontology. And sometimes it is difficult to make
a universal accepted ontology. And a lot of hidden relations are
not modeled at all. This section suggests combining classical
ontology development with data mining for identifying hidden
information and expanding the application area of both
techniques. This gives an improved description of the target
system represented by a database; it can also improve strategy
selection and other forms of decision making.
Data mining techniques to automatically discover and extract
information from Web documents and services [Kosala &
Blockeel, 00, Srivastava et al., 00, Zhong, 01]. Zhong et al.
proposed a way of mining peculiar data and peculiarity rules
that can be used for Web-log mining [Zhong, 99]. They also
proposed ways for targeted retailing by mining classification
rules and retail value functions [Yao and Zhong, 01, Zhong et
al., 00]. Data mining is the systematic refining of information
resources on the Web for business intelligence [Hackathorn,
00].
This paper suggests using associative formal concept
analysis as the base tools for data mining (Wolff, 94) and
develops it further using Dynamic causal mining as the
technique for mining the relations. The DCM algorithm was
discovered in 2005 [Pham et al., 2005] using only counting
algorithm to integrate with Game theory. It was extended in
2006 [Pham et al., 2006] with delay and feedback analysis, and
was further improved for the analysis in Game theory with
Formal Concept analysis [Wang, 2007]. DCM enables the
generation of dynamic causal rules from data sets by
integrating the concepts of Systems Thinking [Senge et al.,
1994] and System dynamics [Forrester, 1961] with Association
mining [Agrewal et al., 1996]. The algorithm can process data
sets with both categorical and numerical attributes. Compared
with other Association mining algorithms, DCM rule sets are
smaller and more dynamically focused. The pruning is carried
out based on polarities. This reduces the size of the pruned data
set and still maintains the accuracy of the generated rule sets.
The rules extracted can be joined to create dynamic policy,
which can be simulated through software for future decision
making. The rest of this section gives a brief review of
Association mining and System Dynamics.
Association mining was discovered by Agrawal [Agrawal et
al., 1996]. It was further improved in various ways, such as in
speed [Agrawal et al., 1996 and Cheung et al., 1996] and with
parallelism [Zaki et al., 1997] to find interesting associations
and/or correlation relationships among large sets of data items.
It shows attributes value conditions that occur frequently
together in a given dataset. It generates the candidate itemsets
by joining the large itemsets of the previous pass and deleting
those subsets which are small in the previous pass without
considering the transactions in the database. By only
considering large itemsets of the previous pass, the number of
candidate large itemsets is significantly reduced.
Systems thinking is about the interrelated actions which
provide a conceptual framework or a body of knowledge that
makes the pattern clearer [Senge et al., 1994]. It is a
combination of many theories such as soft systems approach
and system theory [Coyle, 1996]. Systems thinking seeks to
explore things as wholes, through patterns of interrelated
actions. System dynamics [Sterman 1994 & Coyle, 1996], is a
tool to visualize and understand such patterns of dynamic
complexity, which is build up from a set of system archetypes
based on principles in System thinking [Sterman, 2000]. System
dynamics visualizes complex systems through causal loop
diagrams. A causal loop diagram consists of a few basic shapes
that together describe the action modeled.
System dynamics addresses two types of behavior,
sympathetic and antipathetic [Pham et al., 2005]. Sympathetic
behavior indicates an initial quantity of a target attributes starts
to grow, and the rate of growth increases. Antipathetic behavior
indicates an initial quantity of a target attributes starts either
above or below a goal level and over time moves toward the
goal.
VI. M
INING ALGORITHM
A. Problem Formulation
Let D denote a database which contains a set of n records
with attributes {A1, A2, A3,…, An.}, where each attribute is of
a unique type (sale price, production quantity, inventory
volume, etc). Each attribute is linked to a time stamp t. To apply
DCM, the records are arranged in a temporal sequence (t = 1,
2,…, n). The causality between attributes in D can be identified
by examining the polarities of corresponding changes in
attribute values. Let D
new
be a new data set constructed from D.
A generalized dynamic association rule is an implication of the
form A1 →p A2, where A1
⊂
D, A1
⊂
D, A1∩ A2=
φ
and p
is the polarity.
The implementation of the DCM algorithm must support the
following operations:
(1) To add new attributes.
(2) To maintain a counter for each polarity with respect to every
dynamic value set. While making a pass, one dynamic set is
read at a time and the polarity count of candidates supported by
the dynamic sets is incremented. The counting process must be
very fast as it is the bottleneck of the whole process.
B. Algorithm Description
DCM makes two passes over the data as shown in Figure 3.
In the first pass, the support of individual attributes is counted
and the frequent attributes are determined. The dynamic values
are used for generating new potentially frequent sets and the
actual support of these sets is counted during the pass over the
data. In subsequent passes, the algorithm initializes with
dynamic value sets based on dynamic values found to be
frequent in the previous pass. After the second of the passes,
the causal rules are determined and they become the candidates
for the dynamic policy. In the DCM process, the main goal is to
find the strong dynamic causal rule in order to form a policy. It
also represents a filtering process that prunes away static
attributes, which reduces the size of the data set for further
mining.
Yi Wang
Part 1: – Preprocessing: Removal of the “least” causal
data from database
Part 2: – Mining: Formation of a rule set that covers all
training examples with minimum number of rules
Part 3: – Checking: Check if an attribute pair is self
contradicting (sympathetic and antipathetic at the same
time)
Input: The original database, the values of the pruning
threshold for the neutral, sympathetic and antipathetic
supports.
Output: Dynamic sets
Step 1: Check the nature of the attributes in the original
database (numerical or categorical). Initialize a new
database with dynamic attributes based on the attributes
and time stamps from original database.
Step 2: Initialize a counter for each of the three
polarities.
Step3: Prune away all the dynamic attributes with
supports above the input thresholds.
Step 4. Check weather a rule is self-contradictory (a rule
is both sympathetic and antipathetic).
Step 5. If step 1 returns true then
Retrieve the attribute pair form the preprocessed
database
Step 6. Initialize a counter that includes polarity
combination
Step 7. For the pair of attributes
Count the occurrence of polarity combination with two
records each time.
Prune away the pairs if the counted support is below the
input threshold.
Figure 3. The Steps of DCM
Table I
Pruned results
Single Support
Data set
0.05 0.10 0.15 0.20 0.25 0.30 0.35
Adult
5% 20% 27% 74% 100% 100% 100%
Bank
11% 20% 60% 94% 100% 100% 100%
Cystine
5% 33% 70% 100% 100% 100% 100%
Market basket
6% 10% 50% 86% 100% 100% 100%
Mclosom
1% 13% 38% 72% 90% 100% 100%
ASW
1% 8% 40% 68% 70% 97% 100%
Weka-base
6% 25% 52% 86% 100% 100% 100%
VII. EXPERIMENT
A. Data preparation
The overall aim is to identify hidden dynamic changes. The
original data was given as shown in Table 2. The only data of
interest are the data with changes, for example sale amounts of
a product, the time stamp, etc, The rest of the static data, such as
the weight and the cost of the product can be removed
.
After cleaning the data, the dynamic attributes are found as
shown in Table 3. The dynamic attribute is calculated by
finding the difference between sales amounts in one month and
sales amounts in the previous month.
In the next step, the neutral attributes are pruned. The idea of
pruning is to remove redundant dynamic attributes; thus fewer
sets of attributes are required when generating rules. The first
pruning is based on the single attribute support. In this case, the
single attribute support is defined to be 0.5, which means that if
an attribute with polarity +, -, or 0 occurs in more than half of
total time stamps, it will be pruned. In this case, 429 attributes
remain for the rule generation.
In this experiment, dynamic sets are compared based on a
simultaneous time stamp. Then the support of sympathetic and
antipathetic rules for each dynamic set is calculated. The
support is used as the threshold to eliminate unsatisfactory
dynamic sets and to obtain the rules from the satisfactory sets.
B. Evaluation and Results
The algorithm was run based on the procedures described in
previous sections. Figure 4 shows the plot of sympathetic and
antipathetic support. The x –axis represents the support and the
y-axis represents the number of rules. This database shows that
there are more sympathetic rules than antipathetic rules. The
figure shows that increasing support will lead to exponential
growth of the rules. As the support reaches 0.05 or 5, as it
indicates on the figure, the number of rules is 630. Most of
these rules are redundant and have no meaning due to the low
support. Figure 5 shows the rule plot with support equal to
average value, where the “+support” = “the average of all
Applying Dynamic Causal Mining in Retailing
positive records” and “–support” = “the average of all negative
records”. The number of rules has decreased by applying the
support level.
Table 2 shows the extracted strong rules with support level
equal to average value and support larger than 0.08. There are
only dynamic pairs so there is no need to do the simulation. The
connection can then be put onto concept relational software as
ConExp (Conceptual explorer), and can be represented by a
lattice as shown in Figure 7.
Table II
Result generated by the algorithm
Strong rules
Suppor
t
Sympathetic
{C15276179, F030008}
{J08008008, F060010}
{A04004004, A05005005}
{A05005006, C10251104}
{A04004004, F100020}
0,093
0,089
0,086
0,084
0,082
Antipathetic
{A05005008, C15276179}
{C10251104, F070010}
{A05005008, F030008}
0.092
0.083
0.082
C. Discussion
A priori it is provided some form of causal information, i.e.
suggesting a possible direction of causation between two
attributes, but there is no basis to conclude that the arrow
indicates direct or even indirect causation. The DCM
algorithm, on the other hand, shows causality between
attributes. Thus, where association rule generation techniques
find surface associations, causal inference algorithms identify
the structure underlying such associations.
Each type of relationship generated by the DCM algorithm
provides additional information. The DCM algorithm finds four
kinds of relationships, each of which deepens the user’s
understanding of their target system by constructing the
possible models. For example, A
1
→
+
A
2
provides more
information than A
1
→ A
2
because the latter indicates that A
1
coexists with A
2
. The condition of the rule is not stated
(whether sympathetic or antipathetic). A genuine causality such
as A
1
→
+
A
2
provides useful information because it indicates
that the relationship from A
1
to A
2
is strictly sympathetic causal.
The rules extracted by DCM can be simulated by using
software to model the future behavior. The rules extracted by
association algorithm cannot be simulated.
VIII. C
ONCLUSION
This paper has considered the most fundamental ways to
tackle the problems caused by information overload and
complexity in retailing. The information system would be
available always and everywhere, reacting immediately to any
request for guidance or any change in the situation. It would
constantly be fed with new information thus cause the