Balancing the Analysis of Frequent Patterns
Résumé
A main challenge in pattern mining is to focus the discovery on high-quality patterns. One popular solution is to compute a numerical score on how well each discovered pattern describes the data. The best rating patterns are then the most analyzed by the data expert. In this paper, we evaluate the quality of discovered patterns by anticipating of how user analyzes them. We show that the examination of frequent patterns with the notion of support led to an unbalanced analysis of the dataset. Certain transactions are indeed completely ignored. Hence, we propose the notion of balanced support that weights the transactions to let each of them receive user specified attention. We also develop an algorithm Absolute for calculating these weights leading to evaluate the quality of patterns. Our experiments on frequent itemsets validate its effectiveness and show the relevance of the balanced support.