Alva Erwin Department ofComputing

Raj P. Gopalan, and N.R. AchuthanDepartment of Mathematics and Statistics

Curtin University of Technology Kent St. Bentley Western Australia PAKDD08

Efficient Mining of High Utility Itemsets from Large Datasets

1

OutlineIntroduction

Preliminaries

Method – Compressed Transaction Utility-

Prol

Experiments

Conclusions

2

IntroductionThe goal of frequent itemset mining is to find

items that co-occur in a transaction database above a user given frequency threshold, without considering the quantity or weight such as profit of the items.

Quantity and weight are significant for addressing real world decision problems that require maximizing the utility in an organization.

TwoPhase based on Apriori is suitable for sparse data sets with short patterns, CTU-Mine based on the pattern growth is suitable for dense data.

3

Definition

u(3 4, t1) =$60u(3 4, t3)=$60 u(3 4) = $120 ,

4

DefinitionTransaction Utility :

Transaction weighted Utility:

tu(1) = 80twu(3 4)=$190

TqXDTq

q )tu(T twu(X)

5

Compressed Transaction Utility-Prol

99<min_Utility(129.9)

GlobalItem index

1 2 3 4 5 -

Original item id

5 1 2 4 3 6

Profit 5 10

150

35

25

2

Quantity 60

12

4 5 4 2

TWU 987

964

810

595

422

99

6

Compressed Utility Pattern-

TreeParallel projection of transaction database

7

CUP-tree

Traverse index 1 (110) from 5, 2 (310) from (2,3,4),

3 (195) from 2, and 4 (190)from (3,5)8

ProCUP-treeindex 1 (110) from 5, cause

110<min_Utility(129.9)2 (310) from (2,3,4),3 (195) from 2, and 4

(190)from (3,5)

9

ProCUP-tree

oriUtility*itemQuantity + proUtility*proQuantity = Utility

35*2+25*2=120, 150*1+25*1=175, 10*5+25*3=125

High_Utility_Itemset = (3,2) (3,2,1)

GlobalItem index

1 2 3 4 5

Original item id

5 1 2 4 3

ProItem index

-- 1 2 3 --

Profit 5 10 150

35 25

Quantity 60 12 4 5 4

TWU 987

964

810

595

422

10

Experiments

11

ConclusionCTU-Pro algorithm to mine the complete set of

high utility itemsets from both sparse and relatively dense datasets with short or longer high utility patterns.

The algorithm adapts to large data by constructing parallel subdivisions on disk that can be mined independently.

12