Assessing the statistical significance of association rules

TL;DR

This paper critically examines the statistical significance of association rules, analyzing common measures and techniques to reduce errors, and provides new theoretical insights to improve the identification of truly significant rules.

Contribution

It offers a comprehensive analysis of traditional and alternative measures for association rule significance, identifying limitations and proposing theoretical guidance for better rule discovery.

Findings

Traditional frequency-confidence framework can produce spurious rules.

Certain measures and pruning techniques may lead to type 1 or type 2 errors.

New theoretical results help guide the search for statistically significant rules.

Abstract

An association rule is statistically significant, if it has a small probability to occur by chance. It is well-known that the traditional frequency-confidence framework does not produce statistically significant rules. It can both accept spurious rules (type 1 error) and reject significant rules (type 2 error). The same problem concerns other commonly used interestingness measures and pruning heuristics. In this paper, we inspect the most common measure functions - frequency, confidence, degree of dependence, $\chi^2$, correlation coefficient, and $J$-measure - and redundancy reduction techniques. For each technique, we analyze whether it can make type 1 or type 2 error and the conditions under which the error occurs. In addition, we give new theoretical results which can be use to guide the search for statistically significant association rules.