Efficient Discovery of Correlated Patterns in Transactional Databases Using Items’ Support Intervals

Abstract

Correlated patterns are an important class of regularities that exist in a transactional database. CoMine uses pattern-growth technique to discover the complete set of correlated patterns that satisfy the user-defined minimum support and minimum all-confidence constraints. The technique involves compacting the database into FP-tree, and mining it recursively by building conditional pattern bases (CPB) for each item (or suffix pattern) in FP-tree. The CPB of the suffix pattern in CoMine represents the set of complete prefix paths in FP-tree co-occurring with itself. Thus, CoMine implicitly assumes that the suffix pattern can concatenate with all items in its prefix paths to generate correlated patterns of higher-order. It has been observed that such an assumption can cause performance problems in CoMine. This paper makes an effort to improve the performance of CoMine by introducing a novel concept known as items’ support intervals. The concept says that an item in FP-tree can generate correlated patterns of higher-order by concatenating with only those items in its prefix-paths that have supports within a specific interval. We call the proposed algorithm as CoMine++. Experimental results on various datasets show that CoMine++ can discover high correlated patterns effectively.