An investigation on automatic support thresholds for association rule mining

Abstract

The process of Knowledge Discovery in Database is aimed at extracting useful information from large databases. Among the several steps of Knowledge Discovery in Databases, data mining is the core step. Data mining is the extraction of hidden, predictive information that are implicitly stored in large databases. An association rule will be interesting, if the support and confidence measures of the rules are greater than the user specified support and confidence thresholds. These thresholds play an important role in deciding the quantity and quality of the association rules. The research focuses on assigning automatic support thresholds for the mining of frequent and rare association rules. In view of this, four approaches have been presented in the thesis. The first approach emphasizes the need for levelwise calculation of non uniform support thresholds. In the first approach, a method has been proposed to compute the support thresholds by analyzing the support distribution of items and used for rule mining. Also another method has been proposed to compute support thresholds based on the previous support thresholds. These methods are helpful in extracting large frequent itemsets that are interesting in nature. The first approach is based on the assumption that the items at all levels have same nature and frequency in the database. The need for itemwise support thresholds has been investigated in the second approach. Also, an algorithm based on Confidence Lift Support measure algorithm has been developed to assign support threshold for each item. The algorithm extracts low support but high confidence association rules. The real world datasets consists of items that are of non uniform in nature. Some items appear frequently in dataset and some of them appear rarely. The third approach concentrates on the extraction of rare itemsets. Finally, an approach to deal with the mining of exact association rules has been presented.