Date of Award

2001

Degree Type

Thesis

Degree Name

M.Sc.

Department

Computer Science

First Advisor

Ezeife, C.

Keywords

Computer Science.

Rights

CC BY-NC-ND 4.0

Abstract

Mining association rules among items in a large database have been recognized as one of the most important data mining problems. New transaction insertions and old transaction deletions may lead to previously discovered association rules no longer being interesting, and new interesting association rules may also appear. The process of generating association rules in the updated database using mostly only the updated part of the database and the previous association rules is called incremental association rules maintenance. The most straightforward approach for mining incremental association rules in the updated database starts from scratch to mine the entire database again when update occurs. This approach is very time consuming because it uses the entire database and has to repeat many computations previously done. Some algorithms that utilize the previously discovered frequent patterns have been presented in order to improve the maintenance efficiency by reducing the computation time. However, they still suffer some shortcomings which include: (1) scanning the updated part of the database several times (at each level) to confirm previous large itemsets still large; (2) scanning the entire database several times when some previous small itemsets now become large in the updated part of the database. This thesis proposes two new methods that use the frequent patterns tree (FP-tree) structure to reduce the required number of database scan. One is DB-tree algorithm which stores all the information in a tree structure and requires (1) no scanning of the original database, (2) to only scan the updated transactions once without involving candidate sets generation. Another method is FPUP algorithm, which predicts possible large itemsets for future mining. This FPUP approach also uses FP-tree structure to scan the old database less times than the existing FP algorithms for improved performance. Paper copy at Leddy Library: Theses & Major Papers - Basement, West Bldg. / Call Number: Thesis2001 .S864. Source: Masters Abstracts International, Volume: 40-06, page: 1555. Adviser: Christe Ezeite. Thesis (M.Sc.)--University of Windsor (Canada), 2001.

Share

COinS