The Parallelization of Algorithm based on Horizontal Partition for Association Rule Mining on Large Data Set

DOI : 10.17577/IJERTV9IS120171

Download Full-Text PDF Cite this Publication

Text Only Version

The Parallelization of Algorithm based on Horizontal Partition for Association Rule Mining on Large Data Set

Anil Vasoya Dr. Nitin Koli

Ph. D Scholar Deputy Registrar

Sant. Gadge baba Amravati University, Amravati Sant. Gadge baba Amravati University, Amravati Maharashtra, India Maharashtra, India

Abstract:- Data mining can abstract important facts such as frequent item set from large data set but sometimes it is difficult to achieve all frequent item set if these datasets are split into many clusters when there is a large dataset. In this paper, a detailed comparison has been made for the frequent pattern mining algorithms such as Apriori, ECLAT and FP- Growth based on without partitioning and with horizontal partitioning of dataset by applying actual minimum support as well as with dynamic minimum support. Experimental results shows that after partitioning also we can maintain similar accuracy and have obtained 100% frequent item set in contrast with frequent item set achieved without partitioning. This novel concept is useful to search frequent item set from the large dataset when we applied the concept of partitioning. It also shows that till which extent minimum support should be decreased with respect to number of cluster to achieve maximum accuracy or 100% frequent item set.

Keywords: Association rule Mining, Frequent Items Data Set, Apriori Algorithm, ECLAT, FP- Growth, Minimum support, parallel computing.

  1. INTRODUCTION

    Data mining is concentrate on discovering interesting hidden relationships in large databases. It provides a tool for knowledge discovery to extract unknown and potentially interesting knowledge from data set [1]. One amongst the important techniques utilized in DM is association rules. Association rule is employed to finds interesting patterns among large set of knowledge items, which is an important task of information mining [2, 3].

    Finding frequent itemsets in transaction databases has been proven to be useful in retail business applications [6][11]. Many algorithms are proposed to hunt out frequent itemsets from a very large database. However, there is no published implementation that outperforms every other implementation on every database with every different support [7]. Generally, many implementations are supported these main algorithms: Apriori [2], frequent pattern growth (FP-growth) [4] and ECLAT. The Apriori algorithm discovers the frequent itemsets from an awfully large database through a series of iterations. The Apriori algorithm is required to huge candidate itemsets, compute the support, and prune the candidate itemsets to the frequent itemsets in each iteration. The FP- growth algorithm discovers frequent itemsets with less the time.

    The Apriori algorithm has been look over in numerous ways. One review of the Apriori algorithm is to partition a transaction database into divide partitions DB1, DB2, DB3, , DBn. Partitioning a transaction database may improve the performance of frequent itemsets mining by adjust each partition into main memory for fast access and allowing incremental generation of frequent itemsets. Our implementation is a partition whole dataset in different number of clusters horizontally and generate frequent item set using Apriori algorithm, FP_Growth and ECLAT using parallel computing.

    It observed that once large dataset partition into many cluster then it lost many important frequent item set because of many reason like customer buying pattern, sessional product requirement, seller discount offer etc. Although many algorithms have been proposed, the execution accuracy of frequent pattern mining is still a challenge to the researchers due to the data outburst. Later section of this paper we will see efficient parallel computing approach using dynamic minimum support.

    The rest of this paper is organized as follows:

    • Related work is presented in section II.

    • Proposed methodology is presented in section III.

    • Results and Discussion are presented in Section IV.

    • Conclusions and Future work are discussed in Section V.

  2. LITERATURE REVIEW

    In order to improve the performance of association rule mining, many researchers tried to distribute dataset into the mining computation over more than one processor or a traditional computer have multicore for executing a task. One way of increasing the computational speed is by using all core within a single computer.

    The most famous is the Apriori algorithm which has been brought in 1993 by Agrawal et al. [1] which uses association rule mining[2][3][5] [6][8][9][10]. Association rules are usually required to satisfy a user-specified minimum support and a user-specified minimum confidence at the same time. Association rule generation is usually split up into two separate steps:

    1. Minimum support (threshold) is applied to find all frequent item-sets in a database. 2. These frequent item-sets and the minimum confidence constraints are used to form rules. Advantage of this algorithm, it is easy to find frequent item sets if database is small

    but it has two deadly bottlenecks. First, It needs great I/O load when frequently scans database and Second, It may produce overfull candidates of frequent item-sets.

    Another accomplishment in the development of association rule mining and frequent pattern mining is FPGrowth Algorithm which overcomes the two deficiencies of the Apriori Algorithm . Efficiency of FP-Growth is based on three salient features: (i) A divide-and-conquer approach is used to extract small patterns by decomposing the mining problem into a set of smaller problems in conditional databases, which consequently reduces the search space (ii) FP-Growth algorithm avoid the complex Candidate Itemset generation process for a large number of candidate Itemsets, and (3) To avoid expensive and repetitive database scan, database is compressed in a highly summarized, much smaller data structure called FP tree [4].

    Equivalence class clustering and bottom up lattice traversal (ECLAT) [7] with vertical data format uses intersection of transaction ids list for generating candidate Itemset. Each item is stored with its list of Transaction ids instead of mentioning transaction ids with list of items This algorithm uses breath first search (BFS) and the generation of candidate itemset is needed. Equivalence class clustering and bottom up lattice traversal (ECLAT) algorithm accommodates Depth First Search approach and requires the generation of candidate itemset.

    Proposed a system to reduce the general items and items to be assembled and consequently improve sales in supermarkets. The approach is particularly suitable for analyzing customer performance in relation to purchases of goods and improving the point of sale in supermarkets. An example is customer buying habits that use the Belgian shopping market database, where purchase data, prices and price are used to determine customer behavior and display related information to support sales. In results it has been shown that the time and memory required for the generation of raw materials is reduced by the generation of well-made rules [11].

    Develop a system which proposed for frequent item set mining, where author developed parallel analytical model by integrating R on hadoop [12]. It shows that as the frequent item set is important for inference engine and for analysis it is important to resolve the problems which arise from large dataset. The proposed system takes data into a preprocessing optimizer and it runs it on multistage algorithm. it uses Map reduce application along with som association rules. It showed that a same list is used for operations of mapreduce and is handled with multiple algorithms to give minimized output. One of the drawbacks can be that it doesnt support cyclic flow of data or increamental data.

    Proposed inverted matrix which is a new disk based Parallel association rule mining with minimum inter-processor communication [13]. Apriori algorithm which works on a shared nothing algorithm is not scalable. The inverted matrix has two phases where a global frequent pattern is achieved where the transactional dataset is converted into inverted matrix. The dataset is replicated among many parallel nodes using different support level of association rules, by doing so a global frequent pattern can be generated without communication between node and the load is balanced. The inverted matrix uses combines both horizontal layout and vertical layout and makes the best of both the approaches. The inverted matrix is based on COFI tress. The results of this algorithm acquired 8 nodes with 16 clusters. The paper proposes that we need to have embarrassingly parallel algorithms to store massive side and also to allow random access, and also to allow no communication between nodes.

    Emphasized on the importance of finding frequent item set of large dataset. One of the most effective algorithim to find frequent item set is the Apriori which trims the candidate item set and also finds the support. Hash tress is used in apriori algorithm. The focus of this paper is based on implementing apriroi algorithm using trie structure and analyze its performance of parallel processor. Partitioning helps in quick access and allows the itemsets to be generated incrementally. Trie structure is an effective way to store itemset and to access them . A trie gets created incrementally based on the count of item set . This implementation solved most of the problems of frequent mining faster than most of the implementations but had an issue in load balancing[14].

    K. W. Lin and Y. Luo propose a plan to find fast frequent itemset in cloud computing by utilizing cloud nodes and also prevent the data from getting leaked. The data getting leaked becomes an issues because when we divide the dataset and each processor is distributed with a part to mine , there is an exchange among nodes which increases the load and the database is duplicated to every node. This framework has a kernel cloud which can access the database and has all the information related to the available nodes. As the FP-tree stores data in compressed form as compare to apriori algorithm.

  3. PROPOSED METHODOLOGY

It is difficult to generate frequent item set from large data set in any existing algorithms because existing algorithms create large candidate set hence its impossible to fit into primary memory. This issue can be resolve with the help of parallel computing concept. Divide dataset into many clusters so that it is easy for CPU to generate FIS and it will not face any issue related to memory.

Problem Identified:

As we know, when dataset divides into many clusters or partitioned the dataset it will lost many FIS in process of partition processing. Hence, we proposed a concept dynamic minimum support where it will decrease the minimum support up to certain limit with respect to number of cluster and actual minimum support and next section will show with the help of dynamic support we can achieve almost similar support that we achieved with out clustering.

IV RESULT AND DISCUSSION

Table 1 shows % of decrease in Dynamic minimum support value w.r.t No. of cluster and actual minimum support.

Table 1: % of decrease in Dynamic minimum support value w.r.t No. of cluster and actual minimum support.

No. Of Clusters

% Dynamic Min. Support

1

100.00

3

67.70

5

58.86

7

54.20

9

51.17

11

48.99

13

47.30

15

45.95

17

44.83

21

43.06

23

42.34

25

41.70

27

41.13

31

40.14

33

39.71

35

39.31

37

38.94

39

38.59

Fig 1 shows the deviation of Dynamic Minimum Support over Static Minimum Support with increase in No. of clusters. Dynamic Min. Support is calculated as,

Dynamic Min. Support = Actual Min. Support / Log (No. of Clusters)

Dynamic Min. Support = Actual Min. Support / Log (No. of Clusters)

It indicates the logarithmic decrease in Dynamic Minimum support in order to achieve higher accuracy of FIS in contrast

to lower accuracy of FIS with Static Minimum Support over clustering for Horizontal Approach.

Min. Support ( % )

Min. Support ( % )

100

80

60

40

20

0

1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39

No. Of Clusters

% Dynamic Min. Support % Static Min. Support

Fig. 1: Variation of Dynamic / Static minimum support ( % ) w.r.t No. of cluster.

The Table 2 shows execution time comparison of Normal approach (without clustering) w.r.t Horizontal Approach (with Dynamic / Static Min. support and clustering) for Apriori algorithm over 1 Lakh to 10 Lakh transaction with Minimum support = 0.01 and No. of clusters = 12.

Table 2: Execution Time required by Apriori algorithm with clustering (Dynamic / Static Min. support) and without clustering.

No. Of Transactions

Min. Support

No. Of Clusters

Execution time of Apriori Algorithm (sec)

Normal

Static Support

Dynamic Support

100000

0.1

12

89

53

345

200000

0.1

12

186

101

529

300000

0.1

12

349

153

738

400000

0.1

12

452

207

994

500000

0.1

12

576

246

1362

600000

0.1

12

686

273

1650

700000

0.1

12

728

322

1843

800000

0.1

12

795

379

2062

900000

0.1

12

1091

438

2191

1000000

0.1

12

1201

471

2260

Execution Time (sec)

Execution Time (sec)

2500

2000

1500

1000

500

0

No. Of Transactions

Normal (without clustering) Static Min. Support Dynamic Min. Support

Fig. 2: Performance measure of Apriori algorithm with clustering (Dynamic / Static Min. support) and without clustering.

Fig. 2 shows that Apriori algorithms requires much lesser execution time with static Min. support and in Normal approach in contrast with Dynamic Min. suport. This is due to much higher No. of candidate set generated by Apriori algorithm at lower Dynamic support value.

The Table 3 shows the comparison of No. of Patterns generated in Normal approach (without clustering) w.r.t No. of patterns generated in Horizontal approach (with Dynamic / Static Min. support and clustering) for Apriori algorithm over 1 Lakh to 10 Lakh transaction with Minimum support = 0.01 and No. of clusters = 12.

Table 3: No. of Patterns generated by Apriori algorithm with clustering (Dynamic / Static Min. support) without clustering.

No. Of Transactions

Min.

Support No. Of Clusters

No. Of Patterns (Apriori Algorithm)

Normal

clustering)

100000

0.1

12

115

88

115

200000

0.1

12

115

90

115

300000

0.1

12

116

93

116

400000

0.1

12

116

94

116

500000

0.1

12

117

98

117

600000

0.1

12

116

100

116

700000

0.1

12

118

100

118

800000

0.1

12

118

101

118

900000

0.1

12

119

102

119

1000000

0.1

12

118

102

118

clustering)

100000

0.1

12

115

88

115

200000

0.1

12

115

90

115

300000

0.1

12

116

93

116

400000

0.1

12

116

94

116

500000

0.1

12

117

98

117

600000

0.1

12

116

100

116

700000

0.1

12

118

100

118

800000

0.1

12

118

101

118

900000

0.1

12

119

102

119

1000000

0.1

12

118

102

118

(without

Static Support Dynamic Support

120

100

80

60

40

20

0

No. of Transactions

Normal (without clustering) Static Min. Support Dynamic Min. Support

120

100

80

60

40

20

0

No. of Transactions

Normal (without clustering) Static Min. Support Dynamic Min. Support

No. Of Patterns

No. Of Patterns

Fig. 3: No. of Patterns generated over 1L-10L Transactions. (Apriori Algorithm)

Fig. 3 shows that for Apriori algorithm the accuracy of No. of Patterns generated with Dynamic Min. support is almost equal to 100% while with Static Min. support it drops to approx. 75-85% in contrast with No. of Patterns generated in Normal approach(without clustering).

The Table 4 shows execution time comparison of Normal approach (without clustering) w.r.t Horizontal Approach (with Dynamic / Static Min. support and clustering) for FP-Growth algorithm over 1 Lakh to 10 Lakh transaction with Minimum support = 0.01 and No. of clusters = 12.

Table 4: Execution Time required by FP-Growth algorithm with clustering (Dynamic / Static Min. support) and

without clustering.

No. Of Transactions

Min. Support

No. Of Clusters

Execution time of FP-Growth Algorithm (sec)

Normal

Static Support

Dynamic Support

100000

0.1

12

32

15

193

200000

0.1

12

88

28

41

300000

0.1

12

165

56

63

400000

0.1

12

242

68

72

500000

0.1

12

336

70

86

600000

0.1

12

438

90

116

700000

0.1

12

607

106

131

800000

0.1

12

726

109

136

900000

0.1

12

847

123

139

1000000

0.1

12

995

140

144

Execution Time (sec)

Execution Time (sec)

1000

800

600

400

200

0

No. Of Transactions

Normal (without clustering) Static Min. Support Dynamic Min. Support

Fig. 4: Performance measure of FP-Growth algorithm with clustering (Dynamic / Static Min. support) and without clustering.

Fig. 4 shows that FP-Growth algorithms requires much lesser execution time in Horizontal approach (with Dynamic / Static Min. support and clustering) in contrast with Normal approach. This is due to clustering which results into smaller tree structures in Horizontal approach with lesser branching factor.

The Table 5 shows the comparison of No. of Patterns generated in Normal approach (without clustering) w.r.t No. of patterns generated in Horizontal approach (with Dynamic / Static Min. support and clustering) for FP-Growth algorithm over 1 Lakh to 10 Lakh transaction with Minimum support = 0.01 and No. of clusters = 12.

Table 5: No. of Patterns generated by FP-Growth algorithm with clustering (Dynamic / Static Min. support) and without clustering.

No. Of Transactions

Min. Support

No. Of Clusters

No. Of Patterns (FP-Growth Algorithm)

Normal (without clustering)

Static Support

Dynamic Support

100000

0.1

12

111

84

96

200000

0.1

12

111

87

96

300000

0.1

12

112

90

98

400000

0.1

12

112

91

98

500000

0.1

12

113

95

98

600000

0.1

12

112

97

99

700000

0.1

12

114

97

99

800000

0.1

12

114

98

99

900000

0.1

12

115

99

100

1000000

0.1

12

114

99

99

120

100

80

60

40

20

0

No. Of Transactions

Normal (without clustering) Static Min. Support Dynamic Min. Support

NO. Of Patterns

NO. Of Patterns

120

100

80

60

40

20

0

No. Of Transactions

Normal (without clustering) Static Min. Support Dynamic Min. Support

Fig. 5: No. of Patterns generated over 1L-10L Transactions. (FP-Growth Algorithm)

Fig. 5 shows that for FP-Growth algorithm the accuracy of No. of Patterns generated with Dynamic Min. support is around 85% while with Static Min. support it drops to approx. 75% in contrast with No. of Patterns generated in Normal approach(without clustering).

Table 6: Execution Time required by ECLAT algorithm with clustering (Dynamic / Static Min. support) and without

clustering.

No. Of Transactions

Min. Support

No. Of Clusters

Execution time of ECLAT Algorithm (sec)

Normal

Static Support

Dynamic Support

100000

0.1

12

31

17

55

200000

0.1

12

64

34

63

300000

0.1

12

93

52

80

400000

0.1

12

137

63

92

500000

0.1

12

169

70

106

600000

0.1

12

141

91

109

700000

0.1

12

183

104

114

800000

0.1

12

219

119

118

900000

0.1

12

285

137

120

1000000

0.1

12

355

152

126

The Table 6 shows execution time comparison of Normal approach (without clustering) w.r.t Horizontal Approach (with Dynamic

/ Static Min. support and clustering) for ECLAT algorithm over 1 Lakh to 10 Lakh transaction with Minimum support = 0.01 and No. of clusters = 12.

300

200

100

0

300

200

100

0

No. Of Transactions

No. Of Transactions

Normal (without clustering)

Dynamic Min. Support

Static Min. Support

Normal (without clustering)

Dynamic Min. Support

Static Min. Support

Execution Time (sec)

Execution Time (sec)

Fig. 6: Performance measure of ECLAT algorithm with clustering (Dynamic / Static Min. support) and without clustering.

Fig. 6 shows that ECALT algorithms requires a bit lesser execution time in Horizontal approach with Dynamic / Static Min. support using clustering in contrast with Normal approach.

The Table 7 shows the comparison of No. of Patterns generated in Normal approach (without clustering) w.r.t No. of patterns generated in Horizontal approach (with Dynamic / Static Min. support and clustering) for ECLAT algorithm over 1 Lakh to 10 Lakh transaction with Minimum support = 0.01 and No. of clusters = 12.

Table 7: No. of Patterns generated by ECLAT algorithm with clustering (Dynamic / Static Min. support) and without clustering.

No. Of Transactions

Min. Support

No. Of Clusters

No. Of Patterns (ECLAT Algorithm)

Normal (without clustering)

Static Support

Dynamic Support

100000

0.1

12

115

88

115

200000

0.1

12

115

90

115

300000

0.1

12

116

93

116

400000

0.1

12

116

94

116

500000

0.1

12

117

98

117

600000

0.1

12

116

100

116

700000

0.1

12

118

100

118

800000

0.1

12

118

101

118

900000

0.1

12

119

102

119

1000000

0.1

12

118

102

118

120

100

80

60

40

20

0

No. of Transactions

Normal (without clustering) Static Min. Support Dynamic Min. Support

120

100

80

60

40

20

0

No. of Transactions

Normal (without clustering) Static Min. Support Dynamic Min. Support

No. Of Patterns

No. Of Patterns

Fig. 7: No. of Patterns generated over 1L-10L Transactions. (ECLAT Algorithm)

Fig.7 shows that for ECLAT algorithm the accuracy of No. of Patterns generated with Dynamic Min. support is almost equal to 100% while with Static Min. support it drops to approx. 75-85% in contrast with No. of Patterns generated in Normal approach (without clustering).

V. CONCLUSION

It conclude that in all above algorithms accuracy (FIS) sustained by Dynamic Min. support is almost equal to 100% whereas with Static Min. support it drops to approx. 80% with an increase in No. of clusters. In future execution time can be reduced with the help of additional resources i.e. Multiprocessor environment. Also this research can be extend in future to different category of dataset and with more number of transactions.

REFERENCES

  1. Agrawal, T. Imielinski, and A. Swami. Mining association rules between sets of items in large databases. In Proc. of the ACM SIGMOD Int'l Conf. on Management of Data (A CM SIGMOD '93), Washington, USA, May 1993.

  2. Gosain and M. Bhugra, "A comprehensive survey of association rules on quantitative data in data mining," 2013 IEEE Conference on Information & Communication Technologies, Thuckalay, Tamil Nadu, India, 2013, pp. 1003-1008.

  3. S. P. Aditya, M. Hemanth, C. K. Lakshmikanth and K. R. Suneetha, "Effective algorithm for frequent pattern mining," 2017 International Conference on Energy, Communication, Data Analytics and Soft Computing (ICECDS), Chennai, 2017, pp. 704-708.

  4. Jian Pei, Jiawei Han, "Mining Frequent patterns without candidate generation," in SIGMOD Proceedings of the 2000 ACM SIGMOD international conference on Management of data, New York, NY, USA, 2000, pp. 1-12.

  5. S. Qianxiang and W. Ping, "Association rules mining based on improved PSO algorithm," 2017 2nd IEEE International Conference on Computational Intelligence and Applications (ICCIA), Beijing, 2017, pp. 145-149.

  6. C. Yadav, S. Wang and M. Kumar, "An approach to improve apriori algorithm based on association rule mining," 2013 Fourth International Conference on Computing, Communications and Networking Technologies (ICCCNT), Tiruchengode, 2013, pp. 1-9.

  7. Mohammed J. Zaki, "Scalable Algorithms for Association Mining," IEEE TRANSACTIONS ON KNOWLEDGE AND DATA ENGINEERING, pp. 372-390, 2002.

  8. L.Wang et al., Efficient Mining of Frequent Item Sets on Large Uncertain Databases, IEEE Transactions on Knowledge and Data Engineering, vol.24, no. 12, pp. 2170 2183,Dec. 2012.

  9. Anil Vasoya, Novel Approach to Improve Apriori Algorithm using Transaction Reduction and Clustering Algorithm, International Journal of Applied Information Systems (IJAIS), pp. 37-44, 2014.

  10. Anil Vasoya, Dr. Nitin Koli,Mining of association rules on large database using parallel and distributed computing, 7th International Conference on Communication, Computing and Virtualization 2016, Procidia, Computer science, pp.221 230,2016

  11. Yanbin Ye and Chia-Chu Chiang, "A Parallel Apriori Algorithm for Frequent Itemsets Mining," Fourth International Conference on Software Engineering Research, Management and Applications (SERA'06), Seattle, WA, 2006, pp.87-94.

  12. K. Poorva, H. K. Anushree, K. V. Mahesha, T. R. Pavithra, D. C. Vinutha and S. B. chandini, "Parallel Analytical Model for Frequent Itemset Mining," 2017 International Conference on Current Trends in Computer, Electrical, Electronics and Communication (CTCEEC), Mysore, 2017, pp. 517-519.

  13. M. El-Hajj and O. R. Zaiane, "Parallel association rule mining with minimum inter-processor communication," 14th International Workshop on Database and Expert Systems Applications, 2003. Proceedings., Prague, Czech Republic, 2003, pp. 519-523.

  14. S. Einakian and M. Ghanbari, "Parallel implementation of association rule in data mining," 2006 Proceeding of the Thirty-Eighth Southeastern Symposium on System Theory, Cookeville, TN, 2006, pp. 21-26.

Leave a Reply