You are on page 1of 14

Efficient Tree Structures for High Utility Pattern

Authors:
Chowdhury Farhan Ahmed Syed Khairuzzaman Tanbeer Byeong-Soo Jeong

Source: Springerlink, Studies in Computational Intelligence, Volume: 295/2010, 99-113 Published year:
2010Y

Date: 2010/09/01 Presented by Wayne Chen

Outline

Introduction HUS-Tree HUPMS Experimental results Conclusions

2/14

Introduction (1/3)

High Utility Pattern(HUP) mining

Quantity * Profit = Utility (1) level-wise candidate generation-and-test problem

Whats the problem with HUP mining?

Using Transaction Weighted Utility(TWU) to solve

(2)It needs a large amount of execution time and memory (3)Build once mine many & Interactive mining

by building the data structure only once, several mining operations can be done User can change the threshold and the system doesnt need to rebuild the model.
3/14

Introduction (2/3)

Transaction-Weighted Utilization

twu(de)

= tw(T1) + tw(T8) = 64 + 102 = 166

If twu = minutil = 93.6, twu is High-twu. The downward closure property can be maintained using transaction-weighted utilization

Pattern d, e, and de are all the high-twu.

High twu
4/14

High utility

Introduction (3/3)

Build Once Mine Many

DB

Structure 1

Structure 2

Patterns cover 20% of the total profit

Patterns cover 15% of the total profit


5/14

HUS-Tree (1/2)
Batch Window

6/14

HUS-Tree (2/2)

7/14

HUPMS

All High-twu(candidate) patterns:


a, b, e, d, ab, ad, be, bd, abd

High-twu(candidate) pattern: be:140, e:140 ad:176, bd:228, d:228


8/14

High utility patterns: b:144, ab:134, bd:154, abd:146

Experimental results (1/5)

Environment

OS: Windows XP CPU: Pentium dual-core 2.13 GHz Memory: 2 GB Data:


Synthetic Utility values: muchroom, retail, kosarak Real Utility values: Chain-store

9/14

Experimental results (2/5)

10/14

Experimental results (3/5)

11/14

Experimental results (4/5)

12/14

Experimental results (5/5)

13/14

Conclusions

HUPMS, a novel algorithm for sliding window-based high utility pattern mining over data streams, it can capture the recent change of knowledge in a data stream adaptively by using a novel tree structure It has the build once mine many property and efficient for interactive mining It also saves a huge amount of memory space by keeping the recent information very efficiently in an HUS-tree

14/14

You might also like