Automated support thresholds for rule mining

Abstract

Association rule mining is an important task in data mining which discovers hidden associations between items in the database based on user-specified support and confidence thresholds. To find the relevant associations, an appropriate threshold has to be specified. The support threshold plays a vital role in the quantity and quality of the rules found. The challenge is that one should not miss the rare associations and on the other hand uninteresting associations should not be generated. This paper proposes an approach to obtain the appropriate support thresholds at each level of the level-wise mining approach. It sets the support threshold by analysing the frequency of items and their associations in the database at each level. It uses the central measure of tendency and measure of dispersion to analyse the database and sets the thresholds accordingly. The performance of the proposed approach has been evaluated against multiple sparse and dense datasets. Experimental results show that this approach produces the interesting rules without specifying the user specified support threshold.