Mining top-k regular-frequent itemsets from transactional database

Abstract

Association rule based on support-confident framework is an important task in data mining community. However, the occurrence frequency of a pattern may not be sufficient criterion for mining meaningful patterns. The occurrence behavior can be revealed as an important key in several applications. A pattern is a regular pattern if it regularly occurs in a user-given period (regularity threshold). To mine regular itemsets, a support threshold is used to filter some regular itemsets. However, in practice, it is often difficult for users to provide an appropriate support threshold. Indeed, a too small support threshold could yield a number of regular-frequent itemsets impractically large while a too large threshold could yield very few or no regular-frequent itemsets. Therefore, the use of a support threshold tends to produce a large number of regular-frequent itemsets and it could be better to ask for the number of desired results. Currently, from the deep survey, there is no existing approach permitting users to specify the number of regular-frequent itemsets to be mined. Therefore, a new approach allowing the users to control the number of results (i.e. k regular itemsets with the highest supports) is presented. There are several techniques proposed in this dissertation to mine this kind of itemsets: (i) a single-scan approach, (ii) a best-first search strategy used to prune the search space, (iii) the partitioning and estimation techniques assisting in reducing unnecessary computational costs, and (iv) a new concise representation helping to save runtime and memory. From the performance study, the proposed approaches are efficient and scalable in terms of time and memory for small and large values of desired results on sparse and dense datasets.