On Support Thresholds in Associative Classification

摘要

Associative classification is a well-known technique for structured data classification. Most previous works on associative classification use support based pruning for rule extraction, and usually set the threshold value to 1%. This threshold allows rule extraction to be tractable and on the average yields a good accuracy. We believe that this threshold may be not accurate in some cases, since the class distribution in the dataset is not taken into account. In this paper we investigate the effect of support threshold on classification accuracy. Lower support thresholds are often unfeasible with current extraction algorithms, or may cause the generation of a huge rule set. To observe the effect of varying the support threshold, we first propose a compact form to encode a complete rule set. We then develop a new classifier, named L_G~3, based on the compact form. Taking advantage of the compact form, the classifier can be built also with rather low support rules. We ran a variety of experiments with different support thresholds on datasets from the UCI machine learning database repository. The experiments showed that the optimal accuracy is obtained for variable threshold values, sometime lower than 1%.