INTERACTIVE DATAMINING PROCESS BASED ON
HUMAN-CENTERED
SYSTEM FOR BANKING MARKETING
APPLICATIONS
Olivier Couturier, Engelbert Mephu Nguifo
CRIL CNRS FRE 2499 - Universit
´
e d’artois
rue Jean Souvraz, SP 18, F-62307 Lens Cedex, France
Brigitte Noiret
Caisse d’Epargne du Pas de Calais (CEPDC)
1, place de la r
´
epublique, B.P. 199, F-62304 Lens Cedex, France
Keywords:
Human-Computer Interaction, Knowledge Discovery in Databases (KDD), Association Rules Mining, Bank-
ing Marketing.
Abstract:
Knowledge Discovery in Databases (KDD) is the new hope for marketing due to the increasing collection of
large databases. There is a paradox because the companies must improve the development policy of customer
loyalty by using methods that do not allow to treat large quantities of data. Our current work is the results of
a study that we led on a association rules mining in banking marketing problem. Our first encouraging results
steered our work towards a hierarchical association rules mining, using a user-driven approach rather than an
automatic approach. The user is at the heart of the process, playing a role of evolutionary heuristic. Mining
process is oriented according to intermediate expert’s choices. The final aim of our approach is to use the
advantages of the methods to decrease both number of generated rules and expertise time. This paper presents
the results of our research step for including the user into datamining process.
1 INTRODUCTION
Marketing process is a social and financial mecha-
nism in which people satisfy their needs and desires
by creating and exchanging products or other entities.
The purpose is to establish satisfactory relations be-
tween customers and suppliers in order to preserve
the trade preference. The final aim is to go towards
a network which constitutes the effective capital of
the companies. The sale of new products or main-
tenance products is a necessary step in order to se-
cure customer loyalty. The problem is to anticipate
their choices so as to predict the products which they
might find interesting. There are two categories of
customers. First of all, the customers whom the com-
pany wishes to keep as a customer and secondly, those
from the other banks. For the first case, the term
which is usually used is defensive marketing while
for the second case, the corresponding term is offen-
sive marketing. Our work is oriented towards the first
problem.
Let P , be the population and S, be a sub-
population. The problem is to find the better sub-
population S in P depending on the initial problem.
The implementation of a campaign on launching a
new product can be costly depending on the number
of people who must be contacted. The main idea is to
have a maximum of positive answers in order to re-
duce expenses. It is necessary to be able to offer the
products as well as possible by using the database his-
tory. Indeed, databases contain a significant quantity
of knowledge which is hidden in meaningful masses.
The association rules mining is one of the possible so-
lutions allowing to solve the problem by establishing
logical relations between products. The next section
describes association rules mining.
As the number of large databases increasing, ex-
tracting useful information is a difficult and opened
problem. This is the goal of an active research domain
: Knowledge Discovery in Databases (KDD) (Fayyad
et al., 1996). We focus on association rules mining
(Agrawal et al., 1996) such as “If Antecedent then
Conclusion”. Several works were published which
are based on two main indices to measure the qual-
ity of a rule: support and confidence (Agrawal et al.,
1996). Two main limitations occurs from these works.
104
Couturier O., Mephu Nguifo E. and Noiret B. (2005).
INTERACTIVE DATAMINING PROCESS BASED ON HUMAN-CENTERED SYSTEM FOR BANKING MARKETING APPLICATIONS.
In Proceedings of the Seventh International Conference on Enterprise Information Systems, pages 104-109
DOI: 10.5220/0002550301040109
Copyright
c
SciTePress
First, the support does not allow to extract specific
information valid on a small number of transactions
in the database. Trivial information may already be
known by the user. The calculation of this informa-
tion has a high time cost. This type of approach is an
automatic approach.
Secondly, the number of generated rules is too sig-
nificant and leads to another problem called “Knowl-
edge Mining” (Blanchard et al., 2003). Expertise time
is costly and it is not taken into account in studies.
Indeed, if the support decreases, the number of gen-
erated rules increases. The solution consists in using
quality measures to rate the rules but the experience of
the expert is not taken into account either. In order to
exploit his tacit knowledge, hierarchical association
rules mining can be used, in order to generate knowl-
edge with various levels of granularity (Han and Fu,
1995; Hipp et al., 1998; Srikant and Agrawal, 1997;
Srikant et al., 1997; Tseng, 2001).
In a real-expert datamining process, it is not always
possible to formalize the tacit knowledge of an expert
in order to optimize the rules because he does not al-
ways know which kind of rules he would like to ob-
tain. To solve it, the expert is introduced while the
process is ongoing. Our approach deals with this case.
It is called a user-driven approach and it gives the
main role of evolutionary heuristic to the user (Kuntz
et al., 2000). The main idea of our approach is to use
the advantages of the two methods to decrease both
expertise time and number of generated rules. The hi-
erarchical association rules mining is used to decrease
the number of generated rules and the user-driven ap-
proach is used to reduce trivial rules.
Section 2 recalls association rules mining and gives
a brief overview of both automatic and user-driven ap-
proaches. Section 3 describes our hybrid approach,
and experimental results are presented in section 4.
2 ASSOCIATION RULES MINING
Association rules mining (Agrawal et al., 1996) can
be divided into two subproblems: the generation of
the frequent itemsets lattice and the generation of as-
sociation rules. The complexity of the first subprob-
lem is exponential. Let |I| = m the number of
items, the search space to enumerate all possible fre-
quent itemsets is equal to 2
m
, and so exponential in m
(Agrawal et al., 1993).
2.1 Problem
Let I = {a
1
, a
2
, ..., a
m
} be a set of items, and let
T = {t
1
, t
2
, ..., t
n
} be a set of transactions etablish-
ing the database, where every transaction t
i
is com-
posed of a subset X I of items where each trans-
action has a unique identifier, called T ID. A set of
items X I is called itemset. A subset of items
X I is called a k-itemset. A transaction t
i
contains
an itemset X in I, if X t
i
. The support of an item-
set X, noted σ(X), is the percentage of transactions
contained in T in which X is a subset :
support(X) =
|{t T |X t}|
|t T |
An itemset is frequent if the support σ(X)
minsup, where minsup is the user-specified minimum
support. An association rule is an implication such
as X
1
7→ X
2
, where X
1
and X
2
are itemsets with
X
1
, X
2
I and X
1
X
2
= . The support of an
association rule r: X
1
7→ X
2
is equal to the support
of the union of itemsets which establish it (Agrawal
et al., 1993) :
support(r) = support(X
1
X
2
)
The confidence of an association rule is the condi-
tional probability that the transaction contains X
2
knowing X
1
:
confidence(r) =
support(r )
support(X
1
)
Rules that satisfy both a minimum support threshold
and a minimum confidence threshold are called strong
rules. The following subsection presents automatic
and user-driven approaches.
2.2 Automatic approach vs
user-driven approach
Association rules mining is an automatic task where
the user appears at the beginning and at the end within
the process. First of all, he determines the support
and the confidence of the algorithm. Once the mining
ended, he rates the results obtained. The main prob-
lem of this method is the large number of generated
rules (Kuntz et al., 2000). To solve it, the user-driven
approach applied to association rules mining was pro-
posed. The last two stages of KDD: datamining and
post-processing was grouped. The user is at the heart
of mining and he can drive it throughout the process
(Kuntz et al., 2000). This is the main difference with
automatic approach. The main problem of this ap-
proach is to present comprehensive and fast results so
that the user does not waste his time analyzing cur-
rent results. Various works on rules filtering by qual-
ity criteria were proposed (Ohsaki et al., 2004) but the
experience of the user is not exploited during all the
KDD process. We present related works in the next
subsection.
INTERACTIVE DATAMINING PROCESS BASED ON HUMAN-CENTERED SYSTEM FOR BANKING
MARKETING APPLICATIONS
105
2.3 Related works
All techniques currently developped have a common
purpose which is to discover association rules in data-
bases. One of the major problems of association rules
mining is the large number of patterns which are gen-
erated. It is difficult for the expert to identify those as-
sociation rules that are interesting for him. To help the
expert to choose among these patterns, several works
were proposed on association rules filtering accord-
ing to their interestingness (Liu et al., 1996). Other
works are proposed on hierarchical association rules
mining. Let I and T as previously presented, and
let G, a taxonomy or hierarchical tree. A taxonomy
is a directed acyclic graph (DAG) on the items in I
where items are the leaves and where edges are in-
heritance relation. This approach allows to generate
multi-level association rules (Han and Fu, 1995; Hipp
et al., 1998; Srikant and Agrawal, 1997; Srikant et al.,
1997; Tseng, 2001).
In these works, there are two different kind of rel-
evance measures. The first are called objective mea-
sures (Silberschatz and Tuzhilin, 1995). These mea-
sures are data-oriented (Bayardo and Agrawal, 1999;
Tan et al., 2002; Hilderman and Hamilton, 2001).
Several of these works are efficient to discover the
best rules or to estimate the best rules (Morimoto
et al., 1998) thanks to one of these objective measures
but they are limited. Indeed, there are no measures
which are able to treat efficiently random problems.
They are successful in particular contexts. Further-
more, they do not allow yet to rate patterns quality
(Liu et al., 1996). A comparison of these measures
was done in (Ohsaki et al., 2004).
The second measures are called subjective mea-
sures. These second measures are user-oriented (Pad-
manabhan and Tuzhilin, 1998; Liu et al., 1996). The
expert’s tacit knowledge, in a research domain, are
taken into account. Contrary to objective measures
which are numerous, subjective measures is a recent
research domain. We can count a dozen measures
and split them into three sub-groups: unexpected
(Matheus et al., 1996; Klemettinen et al., 1994; Sil-
berschatz and Tuzhilin, 1996; Liu et al., 1996; Liu
et al., 1997; Padmanabhan and Tuzhilin, 1999; Liu
et al., 2000; Shekar and Natarajan, 2002), actionabil-
ity (Klemettinen et al., 1994; Piatetsky-Shapiro and
Matheus, 1994; Freitas, 1999) and anticipation (Rod-
dick and Rice, 2001). These different measures ex-
ploit the knowledge of the expert to confront them
with the results of datamining. The aim is to highlight
useful and more or less unexpected knowledge. In this
paper, we focus on this second family of measures.
The conception of subjective measure is a difficult
task for several reasons. First, experts have not com-
mon interests for a common research domain. Sec-
ondly, given a database, and a knowledge set, differ-
ent experts may be interested in different subsets of
this explicit knowledge. Finally, the expert’s conclu-
sions can vary depending on time and situation. These
various points show how complex the problem is. We
present in the next section, our new method based on
a combination of both user-driven and hierarchical as-
sociation rules mining approaches.
3 HYBRID ALGORITHM FOR
ASSOCIATION RULES MINING
3.1 Algorithm details
First of all, item taxonomy must be created in coop-
eration with the expert of the domain. The aim is to
develop the taxonomy by grouping items possessing
common roots. The rule will be specialized at the fol-
lowing level according to the expert’s choices. The
current association rule mining consists in building a
lattice of frequent itemsets. In our example, the lat-
tice is created depending on level 4 of the taxonomy,
that is the lowest level. We are working on taxonomy
level. The starting level is the highest one (cf Fig 1
(a)).
A B C D E F G H
A B C D E F G H
A B C D E F G H
(a) Level 1 (b) Level 2
Figure 1: Lattice generated at levels 1 and 2
But, association rules are not generated with a sin-
gle itemset. Consequently, the starting level will al-
ways be level 2 (cf Fig 1 (b)). Thenceforth, there are
two possibilities. First of all, association rules are not
generated. In this case, mining is restarted from the
next level. Secondly, rules are generated depending
on level 2 and proposed to the expert. He selects rele-
vant rules for him. The taxonomy is pruned according
to this selection. The corresponding itemsets are kept
A B C D E F G H
A B C D A B E F A B G H C D E F C D G H E F G H
A B C D E F A B C D G H A B E F G H C D E F G H
A B C D E F G H
Figure 2: Lattice generated at level 3
ICEIS 2005 - HUMAN-COMPUTER INTERACTION
106
in the taxonomy, and other itemsets and their inherited
items are pruned (Hipp et al., 1998). We consider, for
our example, that for this level, the taxonomy was not
pruned. Mining is then achieved at the next level (cf
Fig 2). As previously, rules are generated at level 3
(cf Fig 2). For instance, the expert selects one rule “If
A B G H 7→ E F ”. The itemset C D does
not appear in this rule, and will then be pruned from
the taxonomy as well as the inherited items C and D
(cf Fig 3) (Hipp et al., 1998).
A B C D E F G H
A B C D
A B
A B
C D
C D
E F G H
E F
E F
G H
G H
Level 1
Level 2
Level 3
Level 4
Figure 3: Item taxonomy
If no relevant aggregated rules are found, mining
is restarted from the last level of the taxonomy, as
classical association rules mining. But, for our ex-
ample, there is no more than 2
8
at the beginning but
2
6
thanks to the taxonomy pruning. The main advan-
tage of this method is the variation of space and time
complexities at the search level. If we increase this
search level, running time are going to decrease. The
useless or coarse association rules will be discovered
from the beginning and they are not developed in the
next levels. The number of generalized rules will be
more or less low. This algorithm uses a fixed sup-
port. This approach allows to generate mono-level
rules. We present our algorithm in the next subsec-
tion.
3.2 Search Hierarchical Association
Rules for Knowledge (SHARK)
SHARK algorithm is described hereafter (see algo-
rithm 1). The Update function is not detailed because
it is a simple prunning of the items for the current
level. The RuleGeneration() function is not detailed
too in this paper because this function doesn’t present
a new idea. It runs Apriori which is a parameter that
can be replaced by any other rule generation algo-
rithm. Starting at level i, the algorithm updates the
set of valid items at this level. Given the minsup,
a search for frequent itemsets is achieved with the set
of selected items. A procedure RuleGeneration is run,
given a confidence thereshold, to generate a rules set.
This rules set is analyzed by the expert which selects
relevant ones. Items which are not present among rel-
evant rules are removed from the items list. The pro-
cedure restarted at the next level i + 1 until the last
level is reached. We have implemented a datamin-
ing platform, Lminer, in which both Apriori (Agrawal
et al., 1996) and SHARK are included. SHARK is an
extension of Apriori algorithm that integrates our hy-
brid approach. We present in the next section experi-
mental results.
Algorithm 1: SHARK Algorithm
function AssociationRuleGenerate(Set of items I,
Database T , Taxonomy G)
Data: Set containing all the items,
Database containing all the transactions,
Item taxonomy
Result: Set of association rules F
ra
begin
Level=1;
while NextLevel==OK and user!=END do
// Increase level
Level++;
// Updating the set of items depending on the
//search level
I Update(I,Level);
// Frequent itemsets search
F
g
FrequentSearch(I,T ,G,minsup);
// Generation of association rules
F
ra
RuleGeneration(F
g
,G,minconf);
// Presentation of rules to the expert
F
raF inal
ChooseRule(F
ra
);
// Pruning taxonomy
PruneTaxonomy(G,F
raF inal
);
// End test
if Level==MaxLevel then
NextLevel==NotOK;
return F
raF inal
;
end
4 RESULTS
Our purpose is to include a hierarchical mining and a
user-driven approach. We don’t want propose a new
vizualisation methodology. We developed an easy-
to-use graphical interface for our experiments. In-
deed, extracting nuggets is very difficult when the rel-
evant information is hidden in a large amount of data.
Various works already exist to help expert analysis
(Klemettinen et al., 1996; Liu et al., 1999). These
two works were completed by several works for rules
exploration (Blanchard et al., 2003; Ben-Yahia and
Mephu-Nguifo, 2004). A set of association rules is
proposed to the expert at each level. He chooses the
rules which he finds relevant just with a mouse click
then he runs the mining to the following level with a
button and so on.
First of all, we generated the taxonomy with the ex-
perts. The example shows a balanced tree (cf Fig 3)
but it is not always balanced. It has no influence on
the algorithm running. When the leaves are not on
the last level of the taxonomy, the algorithm selects
INTERACTIVE DATAMINING PROCESS BASED ON HUMAN-CENTERED SYSTEM FOR BANKING
MARKETING APPLICATIONS
107
all the leaves on the previous levels. Our taxonomy
is composed of 114 nodes and 57 leaves. To esti-
mate the rules, a graphical interface was developed
to hold some specific items. This tool allows to vary
several metrics: absolute or relative support rate, con-
fidence rate, number of rules, number of items of the
antecedent and the conclusion and to select specific
items of the antecedent and the conclusion. Thanks to
this tool, the number of rules was reduced from 2445
to few dozens according to the selection of three ex-
perts working together. It is the reason which explains
the short expertise time. We tested our data with Apri-
ori (automatic approach) and with Shark (user-driven
approach). Our benchmark is composed of 57 items
(the last level of the taxonomy) and 400 000 trans-
actions. We work with a pentium III 700 Mhz with
128 Mo RAM. The metrics minsup and minconf
are fixed by the experts. Our results are presented in
table 1.
Table 1: Experiment results
Nb rules Calcul Sort Expert Select
Apriori 2445 1140s 56 600s 10
Shark (L1)
Shark (L2) 5 1s 2 30s 2
Shark (L3) 546 70s 16 60s 6
Shark (L4) 2113 1080s 36 300s 8
Shark 2113 1151s 36 390s 8
Apriori generates 2445 rules in 1140 seconds. This
time does not include the expertise time. It is nec-
essary to add about 600 seconds for a total of 1740
seconds. Shark algorithm works on more and less ag-
gregated data. At the first level, no rule is generated.
The mining begins on the second level. Shark gener-
ates 5 rules in 1 second and 30 seconds of expertise
time. At the first level, 546 rules generated in 70 sec-
onds and 60 seconds of expertise time. Finally, at the
last level, as Apriori algorithm, 2113 rules are gen-
erated in 1080 seconds and 300 seconds of expertise
time. It is necessary to note that this results vary ac-
cording to the material used.
The total running time of the algorithm is 1151 sec-
onds and 1140 seconds for Apriori. Shark is longer
than Apriori. It is because the experts uses a inter-
active graphical interface. A mining on a level de-
pends on an interaction between the experts and the
interface. As long as he does not click on a button to
calculate the next level, mining is stopped. It is the
reason why the running time of fewer rules are longer
with Shark algorithm. Anyway, our main problem is
not the running time but the expertise time. The to-
tal expertise time with Shark (390s) is shorter than
Apriori (600s). The number of rules with SHARK
(8 rules) is smaller than the number obtained with
Apriori (10 rules). The two rules was pruned accord-
ing to experts’ aims. The confidence of these two
rules were very low, and the experts qualified them
as noise among the initial set of selected rules. Fur-
thermore in our experiments, the experts was able to
obtain a gain of 35% of the expertise time when using
SHARK, while the number of selected rules remains
almost significantly identical. However, this obser-
vation should be validated on different other applica-
tions and with different experts. In fact, the loss of
rules during the whole process had to be minimized
to increase the efficiency of the method. And this is
also linked to expert’s choices at each level.
5 CONCLUSIONS AND FUTURE
WORKS
The user-driven approach allows to moderate two
problems. On the one hand, the limitations of expert
reflexion and on the other hand, the problem of data
volume which does not allow to obtain quickly rele-
vant results by automatic approach (Silberschatz and
Tuzhilin, 1996). In this paper, we have presented our
work about a hybrid method based on a hierarchical
association rules mining and a user-driven approach.
The main idea is to decrease the expertise time and
the number of generated rules. Hierarchical associ-
ation rule mining is used to decrease expertise time.
Mining is performed level by level and is oriented by
expert’s choices. The direct consequence is that the
number of generated rules is small and these rules are
better targeted because the expert directs the mining
during all the process.
The experiments show that hybridizing the hierar-
chical association rules mining and the user-driven
approach is interesting especially in our applica-
tion case of banking marketing. Indeed, the Shark
methodology allowed to decrease the total expertise
time and the number of rules. A KDD process must
be steered according to the initial purposes of the ex-
pert. Thus, his role is really essential before, after,
but especially during the process. One limitation of
our approach is that the number of generated rules re-
mains nevertheless large. We reduce the problem but
it needs to be improved. Another limitation is that
we define a taxonomy with simple inheritances. We
don’t treat the multiple inheritances as in (Srikant and
Agrawal, 1997).
ACKNOWLEDGMENTS
This work has been partly supported by the “Cen-
tre National de la Recherche Scientifique” (CNRS),
the Association Nationale de la Recherche Scien-
tifique” (ANRT), the “IUT de Lens” and the “Uni-
versit
´
e d’Artois”. A special thanks to the marketing
service of the “Caisse d’Epargne du Pas de Calais”.
ICEIS 2005 - HUMAN-COMPUTER INTERACTION
108
REFERENCES
Agrawal, R., Imielinski, T., and Swami, A. N. (1993). Min-
ing association rules between sets of items in large
databases. In Buneman, P. and Jajodia, S., editors,
Proceedings of the 1993 ACM SIGMOD International
Conference on Management of Data, pages 207–216,
Washington, D.C.
Agrawal, R., Mannila, H., Srikant, R., Toivonen, H., and
Verkamo, A. I. (1996). Fast discovery of association
rules. In Advances in knowledge discovery and data
mining, pages 307–328. American Association for Ar-
tificial Intelligence.
Bayardo, R. J. and Agrawal, R. (1999). Mining the most
interesting rules. In Proc. of KDD’99, pages 145–154,
San Diego, USA.
Ben-Yahia, S. and Mephu-Nguifo, E. (2004). Emulating
a cooperative behavior in a generic association rule
visualization tool. In Proc. of ICTAI’04, Boca Raton,
Florida, USA.
Blanchard, J., Guillet, F., and Briand, H. (2003). Ex-
ploratory visualization for association rule rummag-
ing. In Proceedings of the 4th International Workshop
on Multimedia Data Mining MDM/KDD’03, Wash-
ington.
Fayyad, U., Piatetsky-Shapiro, G., and Smyth, P. (1996).
From data mining to knowledge discovery: An
overview. In Advances in Knowlegde Discovery and
Data Mining, pages 1-34.
Freitas, A. A. (1999). On rule interestingness measures.
Knowledge-Based Systems journal, 12(5-6):309–315.
Han, J. and Fu, Y. (1995). Discovery of multiple-level
association rules from large databases. In Proc. of
VLDB’95, pages 420–431, Z
¨
urich, Switzerland.
Hilderman, R. J. and Hamilton, H. J. (2001). Knowledge
discovery and measures of interest. Kluwer Acae-
demic Publishers.
Hipp, J., Myka, A., Wirth, R., and G
¨
untzer, U. (1998). A
new algorithm for faster mining of generalized asso-
ciation rules. In Proc. of PKDD ’98, pages 74–82,
Nantes, France.
Klemettinen, M., Mannila, H., Ronkainen, P., Toivonen, H.,
and Verkamo, A. I. (1994). Finding interesting rules
from large sets of discovered association rules. In
Proc. of CIKM’94, pages 401–407. ACM Press.
Klemettinen, M., Mannila, H., and Toivonen, H. (1996).
Interactive exploration of discovered knowledge : A
methodology for interaction and usability studies.
Technical report, University of Helsinki.
Kuntz, P., Guillet, F., Lehn, R., and Briand, H. (2000). A
user-driven process for mining association rules. In
Proc. of PKDD’00, pages 160–168, Lyon, France.
Liu, B., Hsu, W., and Chen, S. (1997). Using general im-
pressions to analyze discovered classification rules. In
Proc. of KDD’97, pages 31–36. AAAI Press.
Liu, B., Hsu, W., Chen, S., and Ma, Y. (2000). Analyz-
ing the subjective interestingness of association rules.
IEEE Intelligent systems, 15(5):47–55.
Liu, B., Hsu, W., Mun, L.-F., and Lee, H.-Y. (1996). Find-
ing interesting patterns using user expectations. vol-
ume 11, pages 817–832.
Liu, B., Hsu, W., Wang, K., and Chen, S. (1999). Visually
aided exploration of interesting association rules. In
Proc. of PAKDD’99, pages 380–389, Beijing, China.
Matheus, C. J., Piatetsky-Shapiro, G., and McNeill, D.
(1996). Selecting and reporting what is interesting :
the kefir application to heathcare data. In Advances in
Knowlegde Discovery and Data Mining, pages 495–
516.
Morimoto, Y., Fukuda, T., Matsuzawa, H., Tokuyama, T.,
and Yoda, K. (1998). Algorithms for mining associa-
tion rules for binary segmentations of huge categorical
databases. In Proc. of VLDB’98, pages 380–391, New
York City, New York, USA.
Ohsaki, M., Kitaguchi, S., Okamoto, K., Yokoi, H., and Ya-
maguchi, T. (2004). Evaluation of rule interestingness
measures with a clinical dataset on hepatitis. In Proc.
of PKDD’04, pages 362–373, Pisa.
Padmanabhan, B. and Tuzhilin, A. (1998). A belief-driven
method for discovering unexpected patterns. In Press,
A., editor, Proc. of KDD’98, pages 94–100, New
York, USA.
Padmanabhan, B. and Tuzhilin, A. (1999). Unexpectedness
as a measure of interestingness in knowledge discov-
ery. Decision support system, 27.
Piatetsky-Shapiro, G. and Matheus, C. J. (1994). The inter-
estingness of deviations. In Proc. of KDD’94, pages
25–36, Seattle, USA.
Roddick, J. F. and Rice, S. (2001). What’s interesting about
cricket? on thresholds and anticipation in discovrerd
rules. SIKKDD Explorations, 3(1):1–5.
Shekar, B. and Natarajan, R. (2002). A fuzzy-graph-based
approach to the determination of interestingness of as-
sociation rules. In Proc. of PAKM’02, pages 377–388.
Silberschatz, A. and Tuzhilin, A. (1995). On subjective
measures of interestingness in knowledge discovery.
In Proc. of KDDM’95, pages 275–281, Menlo Park,
Californie, USA.
Silberschatz, A. and Tuzhilin, A. (1996). User-assisted
knowledge discovery : how much should the user be
involved. In Proc. of DMKD’96, Montreal, Canada.
Srikant, R. and Agrawal, R. (1997). Mining generalized as-
sociation rules. Future Generation Computer Systems,
13(2–3):161–180.
Srikant, R., Vu, Q., and Agrawal, R. (1997). Mining associ-
ation rules with item constraints. In Proc. of KDD’97,
pages 67–73.
Tan, P.-N., Kumar, V., and Srivastava, J. (2002). Selecting
the right interestingness measure for association pat-
terns. In Proc. of KDD’02, pages 32–41. ACM Press.
Tseng, S. (2001). Mining association rules with interesting-
ness constraints in large databases. In International
Journal of Fuzzy Systems, 3(2).
INTERACTIVE DATAMINING PROCESS BASED ON HUMAN-CENTERED SYSTEM FOR BANKING
MARKETING APPLICATIONS
109