Classification Using Statistically Significant Rules

Classification based on association rule mining has become a popular technique within the data mining community. However, it has now been emphatically shown that association rules generated solely on the basis of support and confidence are often not statistically significant i.e, the rules generated are artifacts of the particular dataset being mined rather than a relationship inherent in the underlying population (process). This is not surprising because the use of support is driven by its computational and not statistical properties. In this paper we show that mining for statistically significant rules in a classification setting, by “forcing” Fisher’s Exact Test or its continuous approximation to be “anti-monotonic”, results in a) the vast majority of the mined rules being statistically significant by definition, and b) comparable classification performance on balanced datasets and higher performance on imbalanced datasets. All while examining on average only 0.5% of the search space, using 0.4% of the time and finding 0.06% of the number of rules as techniques using the support-confidence framework. We also provide additional evidence against support and confidence – primarily that they are biased in imbalanced datasets. Thus one arrives at an inescapable conclusion: classification based on rule mining by support and confidence thresholds is not necessary, not efficient and perhaps misleading.