Docstoc

IJETTCS-2013-12-07-056.pdf

Document Sample
IJETTCS-2013-12-07-056.pdf Powered By Docstoc
					    International Journal of Emerging Trends & Technology in Computer Science (IJETTCS)
       Web Site: www.ijettcs.org Email: editor@ijettcs.org, editorijettcs@gmail.com
Volume 2, Issue 6, November – December 2013                                    ISSN 2278-6856

               Mining Association Rule by Multilevel
               Relationship Algorithm: An Innovative
                Approach for Cooperative Learning
                                          D.A. Vidhate1, Dr. Parag Kulkarni2
                                  1
                                   Padmashri Dr. Vithalrao Vikhe Patil College of Engineering,
                                             Vilad Ghat, PO MIDC, Ahmednagar

                                                    2
                                                        EKLaT Research Lab, Pune

Abstract: Mining the Data is also known as Discovery of            to changing conditions which is user-friendly by adapting
Knowledge in Databases is to get correlations, trends,             to needs of their individual users, and also can improve
patterns, anomalies from the databases which can help out to       performance over time.
make exact future decisions. However data mining is not the        Association rule mining concept has been applied to
natural. No one can assure that the decision will lead to good     market domain and specific problem has been studied, the
quality results. It only helps experts to understand the data      management of some aspects of a shopping mall, and an
and lead to good decisions.
                                                                   architecture that makes it possible to construct agents
Association Mining is the discovery of relations or
correlations among an item set. An objective is to make rules
                                                                   capable of adapting the association rules has been used.
from given multiple sources of customer database                   A shopping mall is a cluster of independent shops,
transaction. It needs increasingly deepening the knowledge         planned and developed by one or several entities, with a
mining process for finding refined knowledge from data.            common objective. The size, commercial mixture,
Earlier work is on mining association rules at one level.          common services and complementary activities developed
Though mining association rules at various levels is               are all in keeping with their surroundings. A shopping
necessary. Finding of interesting association relationship         mall needs to be managed and, the management includes
among large amount of data will helpful to decision building,      solving incidents or problems in a dynamic environment.
marketing, & business managing.                                    As such, a shopping mall can be seen as a large dynamic
For generating frequent item set we are using Apriori              problem, in which the management required depends on
Algorithm in multiple levels so called Multilevel Relationship
                                                                   the variability of the products, clients, opinions. Our aim
algorithm (MRA). MRA works in first two stages. In third
                                                                   is to develop an open system, capable of incorporating as
stage of MRA uses Bayesian probability to find out the
dependency & relationship among different shops, pattern of        many agents as necessary, agents that can provide useful
sales & generates the rule for learning. This paper gives          services to the clients not only in this shopping centre, but
detail idea about concepts of association mining,                  also in any other environment such as the labor market,
mathematical model development for Multilevel Relationship         educational system, medical care, etc.
algorithm, Implementation & Result Analysis of MRA and             Data Mining refers to extracting knowledge from large
performance comparison of MRA and Apriori algorithm.               quantity of data. Interesting association can be discovered
                                                                   among a large set of data items by Association rule
Keywords:      Apriori Algorithm, Association rule,                mining. The finding of interesting relationship among
Bayesian Probability, Data mining, Multilevel learning             large amount of business transaction records can help in
                                                                   many business decisions making process, such as catalog
1. INTRODUCTION                                                    plan, cross marketing and loss leader analysis [34].
The area of Machine Learning deals with the design of              However, previous work has been focused on mining
programs that can learn rules from data, adapt to                  association rules at a single concept level. There are
changes, and improve performance with experience. In               applications, which need to get associations at multiple
addition to being one of the initial dreams of Computer            concept levels.
Science, Machine Learning has become crucial as                    Real world problem can be expressed in term of
computers are expected to solve increasingly complex               mathematical model and mathematical solutions can be
problems and become more integrated into our daily lives.          found out. Following stages represents the process for
These include identifying faces in images, autonomous              solving the real world problems.
driving in the desert, finding relevant documents in a              Study of basic concepts for mathematical modeling
database, finding patterns in large volumes of scientific           Mathematical Modeling of the system (MRA)
data, and adjusting internal parameters of systems to               Implementation & Result analysis of MRA
optimize performance. Alternatively methods that take              The focus was on working on mathematical model
labeled training data and then learn appropriate rules             development for multilevel association rule mining.
from the data seem to be the best approach to solve the            Multilevel Apriori algorithm and bayesian probability
problems. Furthermore, it needs a system that can adapt            estimation is not combined in any of the previous work. It

Volume 2, Issue 6 November – December 2013                                                                          Page 130
   International Journal of Emerging Trends & Technology in Computer Science (IJETTCS)
       Web Site: www.ijettcs.org Email: editor@ijettcs.org, editorijettcs@gmail.com
Volume 2, Issue 6, November – December 2013                                    ISSN 2278-6856

is the novel move towards the mining association rule.         c is percentage in transactions in T containing X which
Efficiency of original Apriori algorithm has been              also
increased due to multilevel architecture.                      contain Y. i.e
                                                               Support (X Y) = P (XY)                          (2.2)
2. ASSOCIATION RULE                                            Confidence (X Y) = P (Y|X)                    (2.3)
Mining association rule is finding the interesting             Rules that satisfy both minimum support threshold
association or correlation relationship among large set of     (min_sup) and a minimum confidence threshold
data items. With massive amount of data continuously           (min_conf) are called strong.
being collected & stored in database, many industries are      Itemset is nothing but set of items. If it contains n item is
becoming interested in mining association rule from their      a n-itemset. The set {shirt-Bombay ding, jeans-levis} is
database. Relationship among the business traction             2 itemset. The occurrence of itemset is the number of
records can help to design catalog, loss leader analysis,      transactions that contain the itemset. This is known as
cross marketing & other business decision making               frequency or support count of the item set. It satisfies
process.                                                       lowest amount of support if the occurrences frequency of
The discovery of such association can help retailers to        itemset is greater than or equal to the product of min_sup
develop marketing strategies by gaining insight into           & total no of transactions in T. If an itemset satisfy the
which items are frequently purchased together by               minimum support then it is frequent itemset. Association
customers. Such information can lead to increased sale by      mining has two steps process. In first step, find all
helping retailers to do selective marketing & plan their       frequent item sets. All of these item sets will arise at least
shelf space. Motivating example for association rule           as frequently as a pre-determined minimum support
mining is marker basket analysis.                              count. In second step, generate strong association rules
Market basket analysis can also help retailers to plan         from the frequent item sets and must satisfy lowest
which item to put on sale at reduced price. If customer        amount of support and minimum confidence. The overall
tends to purchase shirt of Bombay ding and jeans of Levis      performance of mining association rule is determined by
together, then having a sale on jeans may encourage the        the first step.
sale of shirt as well as jeans. Buying patterns reflects
which items are frequent associated or purchased               2.1 Apriori Algorithm
together. These patterns represented in the form of            It employs an iterative approach known as a level-wise
association rules. For example, customer who purchase          search, where (k − 1) itemsets are used to explore k item
shirt-Bombay ding also tends to buy jeans Levis at the         sets. First, the set of frequent 1-itemsets is found by
same time is represented in association rule (2.1) below.      scanning the database to collect the count for each item &
          Shirt-Bombay ding  jeans-levis                      collecting those items that satisfy minimum support. The
           [supp=2%, conf=60%]                         (2.1)   outcome is denoted L1. Then L1 is used to find L2, which
Rule support & confidence are two measure rules. They          is then used to find L3, and so on, until no more frequent
respectively reflect the usefulness and certainty of           item sets can be got. Getting of each Lk requires one full
discovered rules. A support of 2% for association rule         scan of D.
means that 2% of all transactions under analysis show          To improve the efficiency of the level-wise generation of
that shirt-Bombay ding and jeans-levis are purchased           frequent itemsets, one takes advantage of the Apriori
together. A confidence of 60% means that 60% of                property: All nonempty subsets of a frequent itemset
customers who purchased shirt-Bombay ding also bought          must also be frequent. This property is based on the
jeans Levis. Typically, association rule are considered        following observation. If an itemset A does not satisfy the
interesting if they satisfy both a minimum support             minimum support threshold, min sup, then A is not
threshold and a minimum assurance threshold. Such              frequent; i.e. P (A) < min sup. If an item B is added to the
threshold can be located by users or area expert.              itemset A, then the resulting itemset AB cannot occur
Let I= {i1, i2, i3…………id} set of all items in dataset          more frequently than A. Therefore AB is not frequent
    T= {t1, t2, t3…....…...tn} set of all transactions         either, that is P (AB) < min sup.
                                                               A two-step process is used to find Lk from Lk−1 , for k ≥ 2:
Each transaction ti contains a subset of items chosen from
I. A transaction tj is said to contain an itemset X if X is
                                                               2.1.1 Join step:
subset of tj.
                                                               To find Lk, a set of candidate k−itemsets is generated by
Association rule is an implication of the form of
                                                               joining Lk−1 with itself. Candidate set is denoted by Ck.
              X Y, where X  I, Y I & X ∩Y = Ф
                                                               Suppose L1 and L2 be item sets in Lk−1 . The notation Li
The rule X Y holds in the transaction set T with
                                                               [j] refers to the jth item in Li. Thus in L1, the last item and
support s, where s is percentage of transactions in T that
                                                               the next to the last item are given respectively by L1 [k −1]
contain X U Y. The rule X Y has confidence c in the
                                                               and      L1 [k −2]. Any two itemsets Lk−1 are joined if their
transaction set T if
                                                               first (k −2) items are in frequent. Then members L1 and
                                                               L2 are joined if

Volume 2, Issue 6 November – December 2013                                                                        Page 131
   International Journal of Emerging Trends & Technology in Computer Science (IJETTCS)
       Web Site: www.ijettcs.org Email: editor@ijettcs.org, editorijettcs@gmail.com
Volume 2, Issue 6, November – December 2013                                    ISSN 2278-6856

(L1[1] = L2 [1]) ^ (L1 [2] = L2 [2]) ^ . . . ^                       Table 2.3: Transaction 1-itemset L1 support 2
(L1[k − 2] = L2[k − 2]) ^ (L1 [k − 1] < L2 [k − 1])                         L1 Itemset      Support count
The condition L1[k − 1] < L2 [k − 1] ensures that no                        SH-A            06
duplicates are created. The outcome of itemset formed by                    SH-B            07
joining L1 and L2 is {L1[1], L1[2]. ………. . L1[k − 2], L1[k −                TSH-P           06
1], L2[k − 1]}                                                              TSH-Q           02
                                                                            J-X             02
2.1.2 Prune step:
Set Ck is a superset of Lk, because although all the           To discover the set of frequent 2-itemsets, L2, the
frequent k-itemsets are included in Ck, its members may        algorithm joins L1 with itself to generate a candidate set
or may not be frequent. One could scan the database to         of 2-itemsets, C2. Note that no candidates are removed
determine the count of each candidate in Ck and eliminate      from C2 during the pruning step since each subset of the
any itemset that does not meet the minimum support             candidates is also frequent.
threshold. This would then give Lk. However, Ck can be
huge, and so this could be very time-consuming.                           Table 2.4: Transaction 2-itemset C2
To eliminate the infrequent itemsets, the Apriori property                  C2 itemset
is used as follows. Any (k−1)-itemset that is not frequent                  (SH-A)-(SH-B)
cannot be a subset of a frequent k-itemset. Hence, if any                   (SH-A)-(TSH-P)
(k−1) itemset of a candidate k-itemset is not in Lk−1, then                 (SH-A)-(TSH-Q)
the candidate cannot be frequent either and so can be                       (SH-A)-(J-X)
removed from Ck. This subset testing can be done quickly                    (SH-B)-(TSH-P)
by maintaining a hash tree of all frequent itemsets. We                     (SH-B)-(TSH-Q)
illustrate the use of the Apriori algorithm for finding                     (SH-B)-(J-X)
frequent itemsets in our transaction database, D.                           (TSH)
                                                                            (TSH)
             Table 2.1: Transaction data set                                (TSH)
             Transaction     Transaction
             ID Items        ID Items
             T1 SH           A,SH                              3. MULTILEVEL RELATIONSHIP
             T2 SH           B,TSH                             ALGORITHM
             T3 SH           B, TSH                            Multilevel Relationship algorithm works in three stages.
             T4 SH           A, SH                             In first two stages it utilizes apriori algorithm for finding
             T5 SH           A, TSH                            out frequent itemsets. Third stage of MRA uses bayesian
             T6 SH           B,TSH                             probability to find out the dependency & relationship
             T7 SH           A, TSH                            amongst different shops and generates the rules for
                                                               learning.
             T8 SH           A, SH
                                                               Let the system S be represented as
             T9 SH           A, SH
                                                               S = {I, O, fs |  s }
In the first iteration of the algorithm, each item is a        I = Input Datasets
member of the set of candidate’s 1-itemsets, C1. The           O = Output Patterns
algorithm simply scans all the transactions in order to        O = fs(I)   s
count the number of occurrences of each item.                  fs : I  O be ONTO function
                                                               Objective was to find out pattern of sale from given
     Table 2.2: Transaction 1-itemset C1 with count            dataset of three different shops for particular time period.
           C1 Itemset        Support count                     Input dataset I = {X,Y,Z} such that X = {x1,x2,x3} ,
           SH-A              06                                Y = {y1,y2,y3} and Z = {z1,z2,z3}
           SH-B              07                                Success output O = {P(X0|Y0), P(X0|Z0), P(X1|Y1),
           TSH-P             06                                P(Y1|Z1)……….. }
           TSH-Q             02
           J-X               02                                Multilevel Relationship Algorithm is applied on given
           J-Y               01                                input dataset i.e. I={X,Y,Z} where X = {x1,x2,x3}, Y =
                                                               {y1,y2,y3} and Z = {z1,z2,z3}.
                                                               First stage gives Level 1 association amongst items in the
The set of frequent 1-itemsets, L1, consists of the
                                                               same shop using knowledge base. It is called as local
candidate itemsets satisfying the minimum support count
                                                               frequent itemsets generated in first phase.
of 2. Thus all the candidates in C1, except for {J-Y}, are
                                                               During second stage it uses individual knowledge base
in L1.
                                                               and level 1 association that was generated in stage I from

Volume 2, Issue 6 November – December 2013                                                                       Page 132
   International Journal of Emerging Trends & Technology in Computer Science (IJETTCS)
       Web Site: www.ijettcs.org Email: editor@ijettcs.org, editorijettcs@gmail.com
Volume 2, Issue 6, November – December 2013                                    ISSN 2278-6856

same shops to find out the frequent item sets i.e. x1(0),        5. Similarly the algorithm is applied on Jewelry shop(Y)
x2(3), x3(1)……etc. It is called as global frequent                 & Footwear shop(Z) to determine frequent itemset on
itemsets.                                                          different items.
                                                                 6. First Level output of Apriori algorithm provided
Stage 1:                                                           internal association amongst the items i. e.
At first stage it found out Level 1 association amongst            y1(0)  y1(1),y2(0)  y2(1),y3(0)  y3(1)                 &
items in the same shop i.e. internal relationship between          z1(0)  z1(1), z2(0)  z2(1), z3(0)  z3(1)......etc for
the same item types i. e. x1(0…….n), x2(0………n),                    Jewelry & Footwear shop respectively.
x3(0……..n) within the Cloth shop (X) i.e. O = fs(X)              7. Second level input of Apriori algorithm provided from
Internal relationship between the same item types i. e.            newly generated individual knowledge base, the
y1(0…….n), y2(0………n), y3(0……..n) within the                        frequent item sets i.e. y1(0), y2(3), y3(1), z1(1), z2(5),
Jewelry shop (Y) i.e. O = fs(Y)                                    z3(6)……
Internal relationship between the same item types i. e.          8. It gives with sets of frequent item sets for the Jewelry &
z1(0…….n), z2(0………n), z3(0……..n) within the                        Footwear shop for different items i.e. Fy & Fz.
Footwear shop (Z) i.e. O = fs(Z)                                 9. The context is generated under uncertainty in the form
                                                                   of frequent item sets Fx, Fy & Fz. System constraints
Stage 2:                                                           applied here are sale of items in a day, week, month or
During second stage it uses individual knowledge base              any particular season. This context is refereed as Fi
and level 1 association is generated in stage 1 of same            which is not constant, i.e. it changed seasonably.
shop to find out the frequent item sets i.e. x1(0), x2(3),      10. Hence it is necessary to determine dynamic behavior
x3(1)……etc is called as global frequent itemsets.                  of Fi for particular season.
It gives sets of frequent item sets for the Cloth shop for      11. External        Dependencies         amongst         Items
different items i.e. Fx as O = fs(x1,x2,x3)                        Xi  Yi….Xn  Yn is found with Bayesian
 It gives sets of frequent item sets for the Jewelry shop for      probability.
different items i.e. Fy as O = fs(y1,y2,y3)                     12. New patterns are generated by Bayesian probability
It gives with sets of frequent item sets for the Footwear          though which learning rules could be predicted &
shop for different items i.e. Fz as O = fs(z1,z2,z3)               interpreted.

Stage 3:
It is necessary to determine dynamic behavior of Fi for
                                                                4 ARCHITECTURE OF MRA
particular season. External Dependencies amongst Items
Xi Yi……... Xn Yn has been found with Bayesian
probability. New patterns are generated by Bayesian
probability through which learning rules are predicted &
interpreted.

3.1 Working of Multilevel Relationship Algorithm
Let the sale of Item X at Cloth shop affects sale of item Y
at Jewelry shop and item Z at Footwear.
1. Apriori association mining algorithm is applied on
   each item in cloth shops separately i.e. Jean(X0),                     Figure 1 : MRA Architecture Diagram
   Tshirt(X1), Shirt(X2) and so on from the given large
   item sets. It was applied at two levels / phases in the      Figure 1 shows the flow diagram which depicted the
   same shop.                                                   development of Multilevel Relationship Algorithm
2. After applying Apriori algorithm at first level for          (MRA). Multilevel Relationship algorithm worked in
   different support value it provide with the internal         three stages.
   dependency amongst individual items & generate the           In first two stages it utilized association rule mining
   individual knowledge base i.e. x1(0)  x1(1), x2(0)          algorithm for finding out frequent itemsets. Datasets of
     x2(1), x3(0)  x3(1) …....etc. It is called as local      three shops i.e. Cloth, Jewelry & Footwear were given as
   frequent itemsets generated in first phase.                  an input to the stage I and Level 1 association between
3. At second level Apriori algorithm was applied on             individual items had been found out. Level 1 association
   newly generated individual knowledge base to find out        between individual items was given as an input to stage II
   the frequent item sets i.e. x1(0), x2(3), x3(1)……etc. It     and frequent itemsets had been found out. These frequent
   is called as global frequent itemsets.                       itemsets had generated new sale context. In stage III it
4. It provided with sets of frequent item sets for the Cloth    used bayesian probability to find out the external
   shop for different items i.e. Fx.                            dependency & relationship amongst different shops,
                                                                pattern of sale and generated the rules for cooperative
                                                                learning. The algorithm consists of three sub modules:
Volume 2, Issue 6 November – December 2013                                                                        Page 133
     International Journal of Emerging Trends & Technology in Computer Science (IJETTCS)
       Web Site: www.ijettcs.org Email: editor@ijettcs.org, editorijettcs@gmail.com
Volume 2, Issue 6, November – December 2013                                    ISSN 2278-6856

MRA Stage I, MRA stage II, Interdependency Module                                      Xi  Yi……Xn  Yn is found with Bayesian
                                                                                       probability. New patterns are generated by Bayesian
MRA Stage I:                                                                           probability through which learning rules are predicted &
At first stage it finds Level 1 association amongst items in                           interpreted. Dependency between itemsets of Cloth shop
the same shop i.e.                                                                     (Fx) and Jewelry shop (Fy) is found out as
Internal relationship between the same item types i. e.
x1(0…….n), x2(0………n), x3(0……..n) within the                                                                           P (Y | X ) P ( X )
                                                                                                       P( X | Y ) 
Cloth shop (X) i.e.                                                                                                        P (Y )
O = fs(X) = fstage_I_algorithm_apriori (X)
 O=fstage_I_algorithm_apriori{x1(….n)}={x1(0)  x1(1), x1(3)  x1(2)…}
   O=fstage_I_algorithm_apriori{x2(0…n)}={x2(2)  x2(4),
x2(2)  x2(4)…}
   O=fstage_I_algorithm_apriori{x3(0…n)}={x3(0)  x3(3),
x3(1)  x3(5)…}
                                                                                       Dependency between itemsets of Jewelry shop (Fy) and
Internal relationship between the same item types i. e.                                Footwear shop (Fz) is found out as
y1(0…….n), y2(0………n), y3(0……..n) within the
Jewelry shop (Y)                                                                                                       P ( Z | Y ) P (Y )
                                                                                                       P (Y | Z ) 
O = fs(Y) = fstage_I_algorithm_apriori(Y)                                                                                   P (Z )
                             O=fstage_I_algorithm_apriori{y1(0…n)}={y1(1)  y1(3),
y1(2)  y1(5)…}
    O=                           fstage_I_algorithm_apriori{y2(0…n)}={y2(0)  y2(1),
y2(3)  y2(7)…}
O= fstage_I_algorithm_apriori{y3(0.n)}={y3(2)  y3(3), y3(1)  y3(4)…...}

Internal relationship between the same item types i. e.                                Dependency between itemsets of Footwear shop (Fz) and
z1(0…….n), z2(0………n), z3(0……..n) within the                                            Cloth shop (Fx) is found out as
Footwear shop (Z)                                                                                                    P ( X | Z )P (Z )
                                                                                                       P(Z | X ) 
O = fs(Z) = fstage_I_algorithm_apriori(Z)                                                                                 P(X )
             O=        fstage_I_algorithm_apriori{z1(0…..n)}   = {z1(0)  z1(2),
z1(2)  z1(4)…}
    O=            fstage_I_algorithm_apriori{z2(0….n)}        =   {z2(1)  z2(4),
z2(1)  z2(3)…}
O= fstage_I_algorithm_apriori{z3(0….n)} = {z3(0)  z3(3), z3(2)  z3(5)…}
                                                                                       Three cases are possible for the system to find out
MRA Stage II:                                                                          interdependencies and for the correct prediction of
During second stage it uses individual knowledge base                                  learning rules.
and level 1 association is generated in stage 1 of same
shop to find out the frequent item sets i.e. x1(0), x2(3),                             Case 1: Sale of items in Footwear shop (Z) depends on
x3(1)……etc is called as global frequent itemsets.                                      sale of items in Jewelry shop(Y) and it is in turn depends
It gave sets of frequent item sets for the Cloth shop for                              on sale of items in Cloth shop (X). That means, X is a
different items i.e. Fx as below.                                                      cause of Y and Y is a cause of Z. Result is an increase in
O = fs(x1,x2,x3)                                                                       sale of items in Cloth shop (X) causes increase in sale of
O=fphase_II_algorithm_apriori{x1,x2,x3}={x1(0)  x2(1),x2(3)  x3(                     items in Jewelry shop (Y) which in turn cause increase in
2), x3(0)  x2(2)…}                                                                    sale of items in Footwear shop (Z)[11].
It gives sets of frequent item sets for the Jewelry shop for
different items i.e. Fy as below
                                                                                       Case 2: Sale of items in Footwear shop (Z) & Jewelry
O = fs(y1,y2,y3)
                                                                                       shop (Y) depends on sale of items in Cloth shop (X). That
O=fphase_II_algorithm_apriori{y1,y2,y3}={y1(0)  y2(1),y2(3)  y3(
2), y3(0)  y2(2)……}
                                                                                       means X is a cause of Y and Z. Two child nodes are
                                                                                       independent given the parent. Y & Z are independent
It gives with sets of frequent item sets for the Footwear
                                                                                       given the parent node X. Result is increase in sale of
shop for different items i.e. Fz as below
                                                                                       items in Cloth shop (X) cause increase in sale of items in
O = fs(z1,z2,z3)
O=     fphase_II_algorithm_apriori{z1,z2,z3}          =       {z1(0)  z2(1),
                                                                                       both the shops i.e. Jewelry shop (Y) & Footwear shop (Z)
z2(3)  z3(2), z3(0)  z2(2)…}                                                         [11].

MRA Stage 3:                                                                           Case 3: Sale of items in Footwear shop (Z) depend on
Interdependency by Bayesian Probability                                                sale of items in Cloth shop (X) and Jewelry shop (Y).
It is necessary to determine dynamic behavior of Fi for                                That means X & Y are the causes of Z. A node has two
particular season. External Dependencies amongst Items                                 parents that are independent unless child is given i.e. an

Volume 2, Issue 6 November – December 2013                                                                                                  Page 134
   International Journal of Emerging Trends & Technology in Computer Science (IJETTCS)
       Web Site: www.ijettcs.org Email: editor@ijettcs.org, editorijettcs@gmail.com
Volume 2, Issue 6, November – December 2013                                    ISSN 2278-6856

event may have independent causes. Result is increase in
sale of items in Cloth shop (X) & Jewelry shop (Y) cause
increase in sale of items in the Footwear shop (Z)
provided sale of items in Cloth shop & Jewelry shop does
not depend on each other [11].

5. EXPERIMENTAL RESULTS
The experimental results that have been obtained through
implementing MRA and Apriori algorithm are presented
in this section. Multilevel relationship algorithm applied
for finding the frequent itemset and external dependency
amongst them. It comes up with pattern which can be
further useful for leaning in cooperative system.
Performance of Apriori and MRA has compared for
various factors i.e. strength, support and interdependency.
Graphs show the result comparison between Apriori and
MRA.




                                                                   Figure 8: Percentage Interdependency among the three shops

                                                              Simple Apriori algorithm shows only the frequent
                                                              itemsets in each shop independently. It does not provide
                                                              the internal dependency amongst individual items and
                                                              cannot find out local frequent itemsets. Due to this,
                                                              external dependencies are not found out between different
                                                              shops and become unable to find out the learning rules
Volume 2, Issue 6 November – December 2013                                                                            Page 135
   International Journal of Emerging Trends & Technology in Computer Science (IJETTCS)
       Web Site: www.ijettcs.org Email: editor@ijettcs.org, editorijettcs@gmail.com
Volume 2, Issue 6, November – December 2013                                    ISSN 2278-6856

and pattern of sale. Hence, there is need to develop          [6] Rakesh Agrawal, Tomasz Imielinski and Arun
modified approach which would enable to give internal &            Swami “Database mining: A performance
external dependencies along with the learning rules.               perspective” published in IEEE Transactions on
Multilevel Apriori algorithm and bayesian probability              Knowledge and Data Engineering, 5(6):914 925,
estimation gives the expected results.                             December 1993. Special Issue on Learning and
                                                                   Discovery in Knowledge-Based Databases.
CONCLUSION                                                    [7] Aaron Ceglar & John F. Roddick “Association
The classical apriori algorithm widely used for                    Mining” in ACM Computing Surveys, Vol. 38, No.
association rule mining, this having important factors i.e.        2, Article 5, Publication date: July 2006.
prediction rate and runtime. This system increase the         [8] Baoqing Jiang,WeiWang and Yang Xu “The Math
efficiency of generating association rules based on these          Background of Apriori Algorithm”
analysis and research. The new algorithm Multilevel           [9] Jiawei Han & Micheline Kamber “Data Mining:
Relationship Algorithm is better than the apriori                  Concepts & Techniques” Second Edition, Elsevier
algorithm. Through this algorithm is good to find the              publication.
frequent item sets with minimum support.New pattern are       [10] Pang-Ning Tan, Vipin Kumar & Michael Steinbach
generated by Bayesian probability though which learning            “Introduction to Data Mining” by Pearson Education
rules are predicted and interpreted. Multilevel Apriori            Inc.
algorithm and Bayesian probability estimation is not          [11] Ethem Alpaydin “Introduction to Machine Learning”
combined in any of the previous work. This is novel move           Second Edition, MIT Press by PHI.
towards the mining association rule. Efficiency of original   [12] Tom Mitchell “Machine Learning” McGraw Hill
algorithm has been increased due to multilevel                     International Edition.
architecture. Two passes of algorithm has been performed      [13] Kishor S. Trivedi “Probability & Statistics with
for more accuracy and efficiency. This multilevel                  Reliability, Queuing and Computer Science
approach is especially beneficial when efficiency required         Applications” by PHI.
is important such as in computationally intensive             [14] Liviu Panait Sean Luke “Cooperative Multi-Agent
applications that must be run frequently. This is new              Learning: The State of the Art”, published in Journal
approach applied to the set of data from different shops           of Autonomous Agents and Multi-Agent Systems
for finding frequent item sets and finding external                Volume 11 Issue 3, pp. 387 – 434, 2005.
dependencies amongst them. It comes up with patters           [15] Young-Cheol Choi, Student Member, Hyo-Sung Ahn
which can be further useful for learning in cooperative            “A Survey on Multi-Agent Reinforcement Learning:
algorithms.                                                        Coordination Problems”, IEEE/ASME International
                                                                   Conference on Mechatronics and Embedded Systems
                                                                   and Applications, pp. 81 – 86, 2010.
References                                                    [16] Zahra      Abbasi,      Mohammad         Ali   Abbasi
[1] R. Agrawal, T. Imielinski, and A. Swami “Mining                “Reinforcement Distribution in a Team of
    associations between sets of items in massive                  Cooperative Q-learning Agent”, Proceedings of the
    databases” In Proc. of the ACM SIGMOD Int’l                    9th ACIS International Conference on Software
    Conference on Management of Data, 1993.                        Engineering, Artificial Intelligence, Networking, and
[2] R. Agrawal and R. Srikant “Fast algorithms for                 Parallel/Distributed Computing, IEEE Computer
    mining association rules in large databases” In                Society, pp. 154-160, 2008.
    Proceedings of the Twentieth International                [17] Babak Nadjar Araabi, Sahar Mastoureshgh, & Majid
    Conference on Very Large Databases, pages 487–                 Nili Ahmadabadi “A Study on Expertise of Agents
    499, Santiago, Chile, 1994.                                    and Its Effects on Cooperative Q-Learning” ,IEEE
[3] Mining Frequent Patterns without Candidate                     Transactions on Evolutionary Computation, vol:14,
    Generation - Jiawei Han, Jian Pei, Yiwen Yin                   pp:23-57, 2010.
[4] RakeshAgrawal, Christos Faloutsos, & Arun Swami           [18] Jun-Yuan Tao, De-Sheng Li “Cooperative Strategy
    “Efficient similarity search in sequence databases” In         Learning In Multi-Agent Environment With
    Proc. of the Fourth International Conference on                Continuous State Space”, IEEE International
    Foundations of Data Organization and Algorithms,               Conference on Machine Learning and Cybernetics,
    Chicago, October 1993. Also in Lecture Notes in                pp.2107 – 2111, 2006.
    Computer Science 730, Springer Verlag, 1993, 69-          [19] Dr. Hamid R. Berenji David Vengerov “Learning,
    84.                                                            Cooperation, and Coordination in Multi-Agent
[5] Rakesh Agrawal, SaktiGhosh, Tomasz Imielinski,                 Systems”, in Proceedings of 9th IEEE International
    BalaIyer, and Arun Swami “An interval classifer for            Conference on Fuzzy Systems, 2000.
    database mining applications” Proc. of the VLDB           [20] M.V. Nagendra Prasad & Victor R. Lesser “Learning
    Conference, pages 560-573, Vancouver, British                  Situation-Specific Coordination in Cooperative
    Columbia, Canada, August 1992.                                 Multi-agent Systems” in Journal of Autonomous

Volume 2, Issue 6 November – December 2013                                                                   Page 136
   International Journal of Emerging Trends & Technology in Computer Science (IJETTCS)
       Web Site: www.ijettcs.org Email: editor@ijettcs.org, editorijettcs@gmail.com
Volume 2, Issue 6, November – December 2013                                    ISSN 2278-6856

     Agents and Multi-Agent Systems, Volume 2 Issue 2,         [34] Toshiharu Sugawara & Victor Lesser “Learning to
     pp. 173 – 207, 1999.                                           improve coordinated actions in cooperative
[21] Edmund H Durfee, Victor R Lesser and Daniel D                  distributed problem solving environments”, in
     Corkill “Trends in Cooperative Distributed Problem             Journal of Machine Learning, Volume 33 Issue 2-3,
     Solving”, IEEE Transactions on Knowledge and                   pp.129-153, 1998.
     Data Engineering, 1995.                                   [35] Hamid Berenji & David Vengerov “Advantages of
[22] Sandip Sen & Mahendra Sekaran “Individual                      Cooperation between Reinforcement Learning
     learning of coordination knowledge”, in Journal of             Agents in difficult stochastic problems” in the 9th
     Experimental & Theoretical Artificial Intelligence,            IEEE International Conference on Fuzzy Systems,
     vol.10 issue3, pp. 333–356, 1998.                              vol.2, pp. 871 – 876, 2000.
[23] Ronen Brafman & Moshe Tennenholtz “Learning to
     Coordinate Efficiently: A Model-based Approach”, in
     Journal of Artificial Intelligence Research, Volume
     19 Issue 1, pp. 11-23, 2003.
[24] Michael Kinney & Costas Tsatsoulis “Learning
     Communication Strategies in Multiagent Systems”,
     in Journal of Applied Intelligence, Volume 9 Issue 1,
     pp 71-91, 1998.
[25] Georgios      Chalkiadakis     &     Craig    Boutilier
     “Coordination       in    Multiagent    Reinforcement
     Learning: A Bayesian Approach” in AAMAS '03
     Proceedings of the 2nd International Joint Conference
     on Autonomous agents and multiagent systems, pp.
     709-716, 2003.
[26] Chern Han Yong & Risto Miikkulainen “Coevolution
     of Role-Based Cooperation in Multi-Agent Systems”,
     in technical Report AI07-338, University of Texas at
     Austin, 2007.
[27] Hung H Bui Svetha Venkatesh and Dorota Kieronska
     “A Framework for Coordination and Learning
     among Team of Agents”, in Agents and Multi-Agent
     Systems:       Formalisms,      Methodologies      and
     Applications, Lecture Notes in Artificial Intelligence,
     Volume 1441, 1997.
[28] Jun Huang, N. R. Jennings & John Fox “An Agent
     Architecture for Distributed Medical Care” in
     Lecture Notes in Computer Science, Volume
     890/1995, pp. 219-232, 1995.
[29] Thomas Haynes & Sandip Sen “Adaptation Using
     Cases in Cooperative Groups”, in workshop
     proceedings of Association for the Advancement of
     Artificial Intelligence (AAAI), 1996.
[30] Richardson Ribeiro, André P. Borges and Fabrício
     Enembreck “Interaction Models for Multiagent
     Reinforcement Learning”, in the CIMCA '08
     Proceedings of IEEE International Conference on
     Computational Intelligence for Modeling Control &
     Automation, pp. 464-469, 2008.
[31] Herman Bruyninckx “Bayesian probability” ,Dept. of
     Mechanical Engineering, K.U. Leuven, Belgium,
     November 2002
[32] Chris Westbury “Bayes’ For Beginners” Department
     of Psychology, P220 Biological Sciences Bldg.,
     University of Alberta, Edmonton, AB, T6G 2E9,
     Canada.
[33] Bruno A. Olshausen “Bayesian probability theory”
     March 1, 2004

Volume 2, Issue 6 November – December 2013                                                                  Page 137

				
DOCUMENT INFO
Shared By:
Categories:
Tags:
Stats:
views:0
posted:1/23/2014
language:English
pages:8
Description: International Journal of Emerging Trends & Technology in Computer Science (IJETTCS) Web Site: www.ijettcs.org Email: editor@ijettcs.org, editorijettcs@gmail.com Volume 2, Issue 6, November – December 2013 ISSN 2278-6856, Impact Factor: 2.524 ISRA:JIF
Editor IJETTCS Editor IJETTCS
About