B0950814

IOSR Journal of Computer Engineering (IOSR-JCE)
e-ISSN: 2278-0661, p- ISSN: 2278-8727Volume 9, Issue 5 (Mar. - Apr. 2013), PP 08-14
www.iosrjournals.org

Comparative Evaluation of Association Rule Mining Algorithms
with Frequent Item Sets
Vimal Ghorecha
(Department of MCA, ATMIYA Institute of Technology & Science, India)

Abstract: This paper represents comparative evaluation of different type of algorithms for association rule
mining that works on frequent item sets. Association rule mining between different items in large-scale database
is an important data mining problem. Now a day there is lots of algorithms available for association rule
mining. To perform comparative study of different algorithms various factor considered in this paper like
number of transaction, minimum support and execution time. Comparisons of algorithms are generated based
on experimental data which gives final conclusion.
Keywords – Apriori, Association Rules, Data Mining, Frequent Pattern

I. INTRODUCTION
Many change and many of the developments in association rule mining for the last decade are due to
new algorithms introduced. On the one hand where the main aim of association rule mining is to provide the
better rule and frequent item set to the predict and make decision. This section explains core concept of frequent
itemset and association rule mining. Frequent itemset are those items which occur in transaction frequently. For
example, a person who purchase Health Policy, generally purchases Accident Policy. This both items frequently
appear in database of insurance company.
Association rule mining can be applied in industries like supermarket, insurance company, and online
shopping store and has become essential data mining task which gives different rules for taking future decision.
A very popular and firstly introduced algorithm is Apriori algorithm that has been developed in 1993
by Agrawal et al. [1]. After this frequent itemset mining and association rule mining problems have received
great deal of attention. Within one or two decade many research paper published and many other algorithms
developed based on this Apriori algorithm. An important improvement regarding performance of these
algorithms was made and a new technique introduced called frequent pattern tree (FP-tree) [2]. Also related to
associated mining algorithm developed called FP-growth.
After this introduction, paper is organized as below. Section 2 gives problem definition. After that each
section 3,4,5,6 give brief of different algorithm Apriori, FP-Growth, The Closure Algorithm [3] and The
MaxClosure Algorithm [3] respectively. Section 7 & 8 gives comparative evaluation of these algorithms each
compared with Apriori and comparison of all algorithms together. Finally paper concluded in section 9.

II. DEFINING PROBLEM
Our problem is generally divided into two different sub-problems. The first part of problem is to find
item sets which occurs in database with minimum support; those items sets are known as large item sets or
frequent item sets. The second part of problem is to generate association rules from those large item sets with
the constraints (also there is other constraints) of minimal confidence.
Let I = { i1, i2, i3, i4………. im } be a set of m different items, D is a database, is a set of database
transactions (variable length) where each transaction T contains a set of items i1, i2, i3, i4……….. ik ⊆ I. Each
transaction is associated with an identifier, called TID. A formal definition is [4] Let X is a set which contains
different items. Any transaction T has element X if and only if it the element X is with that transaction T. An
association rule is an implication of the form X  Y, where X, Y ⊆ I and X ∩ Y = Ø. Here X is called the
antecedent and Y is called the consequent of the rule. The rule X  Y satisfies in the dataset with confidence c
if within all transactions that contain X also contains Y with ratio of c%. The rule X  Y has support s in
dataset D, if s% of transaction in dataset D either contain X or Y (X ∪ Y). The selection of association rules is
based on these two values (Other constraints also matters). These are two important measures of rule
importunateness. They can be described by the following equations:
support( X  Y ) = Freq (X ∪ Y) / Total Trns (1)
confidence ( X  Y ) = Freq (X ∪ Y) / Freq(X) (2)
The first rule is {i1, i2, … ,ik-1}  {ik}, by checking the confidence. We can create rules from large
item sets which satisfies minimum support. First we check that rule can be satisfied by our requirement. If rule
is not as per our requirement new rule are generated by removing items from antecedent and inserting this items
to consequent. This new rule again checked for confidence and further rule can be satisfied by our requirements.
www.iosrjournals.org 8 | Page

Comparative Evaluation of Association Rule Mining Algorithms with Frequent Item Sets

The above processes continue until the antecedent does not have any item of say it becomes empty. Our second
problem is very easy so every researcher concentrates on first problem.
The first sub-problem can be further divided into two sub-problems: candidate large item sets
generation process and frequent item sets generation process. We call those item sets whose support exceeds the
support threshold as large or frequent item set.
This requires multiple scan of dataset. Each time for checking support and creating rule based on
frequent itemset we need to scan whole dataset and we can say all transactions.
We already discuss the process of creating rules in this section which divided into two parts [5]. First
one is to identify all frequent item sets. Second generate strong rules which satisfy minimum support and
minimum confidence [6].

III. APRIORI ALGORITHM
Apriori is an algorithm [1] is very popular and previous first algorithm for mining frequent item sets. It
generates candidate itemset and frequent itemset, like at each level are utilized in next level to generate
candidate itemset. As we said earlier it requires multiple database scans, as many as longest as frequent item
sets. Apriori employs an iterative approach known as level-wise search, where k item set are used to explore
(k+1) item sets. There are two steps in each iteration.
The basic Apriori algorithm is plinth of association rule mining classic Apriori algorithm inspires
different researcher how to mine association rules and the market basket analysis example is most popular in
1990s.
L1 = {large 1-itemsets};
for ( k = 2; Lk-1 ≠ Ø; k++ ) do begin
Ck = apriori-gen(Lk-1); //New candidates
forall transactions t ∈ D do begin
Ct = subset(Ck, t);
//Candidates contained in t
forall candidates c ∈ Ct do
c.count++;
end
Lk = { c ∈ Ck | c.count ≥ minsup }
end
Answer = ∪ k Lk;
The first weakness of this algorithm is the generation of a large number of candidate item sets. The
second problem is the number of database passes which is equal to the max length of frequent item set.

IV. THE FP-GROWTH ALGORITHM
The FP-growth [2] algorithm for mining frequent patterns with FP-tree by pattern fragment growth is:
A FP-tree constructed with the above mentioned algorithm; where D – transaction database; s –
minimum support.
Output of this algorithm is complete set of frequent itemset without generating candidate set and
without multiple database scans.

Method:
call FP-growth(FP-tree, null).

Procedure FP-growth(Tree, A)
{
if Tree contains a single path P
then for each combination (denoted as B) of the nodes in the path P do
generate pattern B ∪ A with support=minimum support of nodes in B
else for each ai in the header of the Tree do
{
generate pattern B = ai ∪ A with support = ai.support;
construct B’s conditional pattern base and B’s conditional FP-tree
TreeB;
if TreeB ≠ Ø
then call FP-growth(TreeB, B)
}
}



V. THE CLOSURE ALGORITHM
Candidate = all 1 item sets; KFrequent = empty; Maximal = empty
while (true) do
initialize the elements of matrix M with 0;
count the support for all elements of Candidate
and update their corresponding rows in M;
for all item sets C in Candidate do
if (C is not frequent)
continue to next itemset;
add C to KFrequent;
compute closure of C in cl(C);
compute all extended closures of C and add them to Maximal;
if no extended closures were found
add cl(C) to Maximal;
od
empty Candidate;
generate new candidate item sets in Candidate based on KFrequent;
if Candidate is empty
break;
empty KFrequent;
for all item sets L in Maximal
mark as frequent all item sets from Candidate
those are contained in L;
for all item sets C in Candidate
if C is marked as frequent
add C to KFrequent;
empty Candidate;
generate new candidate item sets in Candidate based on KFrequent;
if Candidate is empty
break;
empty KFrequent;
od
return Maximal
Note that the Closure algorithm requires (n/2 + 1) scans of the database where n are the dimension of the longest
frequent itemset.

VI. THE MAXCLOSURE ALGORITHM
Candidate = all 1-itemsets; Frequent = empty; Maximal = empty;
while (true) do
initialize the elements of matrix M with 0;
count the support for all elements of Candidate
and update their corresponding rows in M;
for all item sets C in Candidate do
if C is not frequent
continue to next itemset;
compute closure of C in cl(C);
for all extended closures xcl(C) of C
if xcl(C) has not been already added to Frequent
add xcl(C) to Frequent;
if no extended closures were found
if cl(C) has not been already added to Maximal
add cl(C) to Maximal;
od
if Frequent is empty break;
empty Candidate;
move into Candidate all elements from Frequent;
od
return Maximal



VII. COMPARATIVE EVALUATION BETWEEN APRIORI, FP-GROWTH, CLOSURE AND
MAXCLOSURE ALGORITHM
Four different association rule mining algorithm implemented in Java and tested based on different
criteria. The platform specification for this test was:
Intel Core-i3-2330M 2.20GHz processor, with 4 GBRAM, Windows 7 64bit. To study the performance
and scalability of the algorithms generated data sets with 10K to 500K transactions, and support factors of 25%
and 33% were used. Any transaction may contain more than one frequent itemset. The number of items in a
transaction may vary, as well as the dimension of a frequent itemset. Also, the number of items in an itemset is
variable. Taking into account these considerations, the generated data sets depend on the number of items in a
transaction, number of items in a frequent itemset, etc.

VII.1. Comparative evaluation between Apriori and FP-Growth
From TABLE-1 and Fig.1 we can tell that FP-Growth algorithm is more efficient than classic Apriori
algorithm. In analysis process we have taken records from 10K to 500K with support of 33% and 100 numbers
of items.
TABLE 1
Apriori Fp-Growth
Records
(Time in Seconds ) (Time in Seconds)
10000 1.358 1.352
50000 6.822 6.651
100000 13.273 13.133
300000 40.906 39.657
500000 99.317 89.314

Fig. 1. Apriori vs. Fp-Growth

When we increase the size of record we can see in the table the results are very efficient related to time.

VII.2. Comparative evaluation between Apriori and Closure

From TABLE-2 and Fig.2 we can tell that Closure algorithm is more efficient than classic Apriori
of items.
TABLE 2
Apriori Closure
Records
10000 1.358 0.722
50000 6.822 3.444
100000 13.273 6.671
300000 40.906 20.907
500000 99.317 68.678



Fig. 2. Apriori vs. Closure


VII.3. Comparative evaluation between Apriori and MaxClosure

From TABLE-3 and Fig.3 we can tell that MaxClosure algorithm is more efficient than classic Apriori
of items.
TABLE 3
Apriori MaxClosure
Records
10000 1.358 0.677
50000 6.822 3.340
100000 13.273 6.563
300000 40.906 20.141
500000 99.317 65.288

Fig. 3. Apriori vs. MaxClosure




VIII. COMPARATIVE EVALUATION BETWEEN ALL ALGORITHMS TOGETHER.

TABLE 4
Apriori
Fp-Growth Closure MaxClosure
Records (Time in Seconds
(Time in Seconds) (Time in Seconds) (Time in Seconds)
)
10000 1.358 1.352 0.722 0.677
50000 6.822 6.651 3.444 3.340
100000 13.273 13.133 6.671 6.563
300000 40.906 39.657 20.907 20.141
500000 99.317 89.314 68.678 65.288

TABLE 5
Apriori MaxClosure
Fp-Growth Closure
Records (Time in Seconds (Time in
(Time in Seconds) (Time in Seconds)
) Seconds)
10000 1.907 1.544 0.803 0.760
50000 7.387 7.411 3.885 3.737
100000 14.822 14.367 7.480 7.270
300000 45.199 44.564 23.461 22.004
500000 75.558 73.640 38.360 37.487

For the comparative study of Classical Apriori, FP-Growth, Closure and MaxClosure Algorithm, we
have taken a database of different transaction form 10K to 500K with different number of items 100 and 200.
In this analytical process we considered 500 transactions to generate the frequent pattern with the
support count 25% and 33%. We have repeated the same process by increasing the transaction.
Here we first compare the result of all approaches with Classical Apriori Algorithm because all
approaches are based on Classical Apriori. After that we compare all approaches to find out the best. After the
experiment on all approaches, we have designed a graph and summarized a result in the following table.



IX. CONCLUSION
From above experimental data we can conclude that the MaxClosure algorithm is better than all
algorithms. We can also conclude that FP-Growth is better than Apriori, Closure is better than FP-Growth and
MaxClosure is better than Closure Algorithm. Here we just consider time as factor. If we consider other factor
other than time the result may be vary from factor to factor. Performance of all algorithms is not affected by
support factor as well as number of items.
If we closely monitor Table – 5 we can say that time taken by Apriori algorithm when records is 500K
and Support is 25%, is less than time taken by Apriori when records is 500K and Support is 33%. So based on
data we can conclude that classical Apriori algorithm affected by Support Factor.
If we closely monitor Figure – 5 we can say that time taken by Closure and MaxClosure when records
are 500K is same as time taken by Apriori and FP-Growth when records is 300K. So based on data we can
conclude that both Closure algorithms take almost half of the time taken by Apriori and FP-Growth.

REFERENCES
[1] R. Agrawal, R. Srikant. “Fast algorithms for mining association rules in large databases” Proc. of 20th Int’l conf. on VLDB: 487-499,
1994.
[2] J. Han, J. Pei, Y. Yin. “Mining Frequent Patterns without Candidate Generation” Proc. of ACM-SIGMOD, 2000.
[3] D. Cristofor, L. Cristofor, and D. Simovici. Galois connection and data mining. Journal of Universal Computer Science, 6(1):60–73,
2000.
[4] R. Győrödi, C. Győrödi. “Architectures of Data Mining Systems”. Proc. Of Oradea EMES’02: 141-146, Oradea, Romania, 2002.
[5] M. H. Dunham. “Data Mining. Introductory and Advanced Topics”. Prentice Hall, 2003, ISBN 0 -13-088892-3.
[6] J. Han, M. Kamber, “Data Mining Concepts and Techniques”, Morgan Kaufmann Publishers, San Francisco, USA, 2001, ISBN
1558604898.
[7] R.J. Bayardo, Jr. “Efficiently mining long patterns from databases”. In L.M. Haas and A. Tiwary, editors, Proceedings of the 1998
ACM SIGMOD International Conference on Management of Data, volume 27(2) of SIGMOD Record, pages 85–93. ACM Press,
1998.
[8] R. J. Bayardo. “Efficiently mining long patterns from databases”. In SIGMOD'98, pp. 85-93.
[9] G. Dong and J. Li. “Efficient mining of emerging patterns: Discovering trends and differences”. In KDD'99, pp. 43-52.
[10] Zaki, M.J. and Hsiao, C.J. "CHARM: An efficient algorithm for closed itemset mining". In Proc. SIAM Int. Conf. Data Mining,
Arlington, VA, pp. 457–473, 2002.


B0950814

More Related Content

What's hot (20)

Viewers also liked (20)

Similar to B0950814 (20)

More from IOSR Journals (20)

B0950814