Comparative performance of interestingness measures to identify redundant and non-informative rules from web usage data

Abstract

Association rules are used to predict frequent web user behaviors from web usage data. These rules are formed using frequent items. The number of association rules increases as the number of frequent items increases and produces several redundant and non-informative rules. In this paper, five interestingness measures, including cosine, lift, leverage, confidence, and conviction with a constant value of support are compared based on the number of redundant and non-informative rules that they produce. Redundant and non-informative rules are a subset of rules present in the top generated rules. The experimental results suggested that leverage produced the least number of redundant rules in the top rules but also produced the least informative rules among all measures. Lift showed the highest number of redundant rules but the most informative rules among all the measures.