Find Interesting Frequent Patterns And Subspaces Using Length-Decreasing Support

Frequent patterns are itemsets, subsequences, or substructures that appear in a dataset with frequency no less than a user-specified threshold. Frequent patterns mining play an important role in mining association rules, correlation rules, and other interesting relations among data, and could be used in data indexing, categorization, clustering and other mining. Therefore, frequent pattern mining is a critical mining and has become a focused theme in data mining.In this thesis, we study on how to find interesting frequent patterns and apply them in subspace selection. We propose a new algorithm which mines LDS-closed frequent itmsets and create a novel subspace quality measure on categorical datasets using LDS-frequent itemsets. Main contributions in the thesis include:[1] A vertical mining algorithm called LDS_CLOSED has been proposed, which could find LDS-closed frequent itemsets very efficiently.[2] Two novel methods, Invalid Prefix Pruning and Pruning based on SVE property, have been put forward to efficiently prune searching space.[3] Experimental results on real datasets show that, LDS_CLOSED not only gets more concise pattern set, but also performs more efficiently than the algorithms that mine closed frequent itemsets.[4] A novel subspace quality measure on categorical datasets are proposed, which require no parameters hard to predict. So the new measure show "non-supervised learning" better.[5] The relationship between subspace quality and LDS-frequent itemsets has been built, which explains why the itemsets not satisfying length-decreasing support constraints could be pruned.