Predicate based association rules mining with new interestingness measure

Abstract

Association Rule Mining (ARM) is one of the fundamental components in the field of data mining that discovers frequent itemsets and interesting relationships for predicting the associative and correlative behaviours for new data. However, traditional ARM techniques are based on support-confidence that discovers interesting association rules (ARs) using predefined minimum support (minsupp) and minimum confidence (minconf) threshold. In addition, traditional AR techniques only consider frequent items while ignoring rare ones. Thus, a new parameter-less predicated based ARM technique was proposed to address these limitations, which was enhanced to handle the frequent and rare items at the same time. Furthermore, a new interestingness measure, called g measure, was developed to select only highly interesting rules. In this proposed technique, interesting combinations were firstly selected by considering both the frequent and the rare items from a dataset. They were then mapped to the pseudo implications using predefined logical conditions. Later, inference rules were used to validate the pseudo-implications to discover rules within the set of mapped pseudo-implications. The resultant set of interesting rules was then referred to as the predicate based association rules. Zoo, breast cancer, and car evaluation datasets were used for conducting experiments. The results of the experiments were evaluated by its comparison with various classification techniques, traditional ARM technique and the coherent rule mining technique. The predicate-based rule mining approach gained an accuracy of 93.33%. In addition, the results of the g measure were compared with a state-of-the-art interestingness measure developed for a coherent rule mining technique called the h value. Predicate rules were discovered with an average confidence value of 0.754 for the zoo dataset and 0.949 for the breast cancer dataset, while the average confidence of the predicate rules found from the car evaluation dataset was 0.582. Results of this study showed that a set of interesting and highly reliable rules were discovered, including frequent, rare and negative association rules that have a higher confidence value. This research resulted in designing a methodology in rule mining which does not rely on the minsupp and minconf threshold. Also, a complete set of association rules are discovered by the proposed technique. Finally, the interestingness measure property for the selection of combinations from datasets makes it possible to reduce the exponential searching of the rules.