Applying multiple minimum support on classification based association

Abstract

Association rule mining is one of the most important algorithms in data mining for Ecommerce. It discovers associations between items in the data that satisfy user-specified thresholds including minimum support and minimum confidence. The application of onesize-fit-all minimum support in the algorithm, however, cannot cover all products because different items have different purchasing frequencies, leading to the loss of many rare items in the output. A solution to this problem is to use multiple minimum supports for items. Yet, this approach has only been studied on Apriori and FP-Growth, which are the most popular algorithms of association rules mining. Our paper applies this approach on the rule generator of Classification Based on Association (CBA-RG), an associative classification algorithm developed based on Apriori, and proposes an improved model with multiple minimum supports called MS-CBA. For classifying the data, we apply ABC classification. In the evaluation, our system achieves fast and stable performance, high efficiency for all data sizes, and outperforms the classic CBA-RG in the number of rare items covered, the number of ruleitems being generated and computational time.