On optimal rule discovery

Abstract

In machine learning and data mining, heuristic and association rules are two dominant schemes for rule discovery. Heuristic rule discovery usually produces a small set of accurate rules, but fails to find many globally optimal rules. Association rule discovery generates all rules satisfying some constraints, but yields too many rules and is infeasible when the minimum support is small. Here we present a unified framework for the discovery of a family of optimal rule sets, and characterise the relationships with other rule discovery schemes such as non-redundant association rule discovery. We theoretically and empirically show that optimal rule discovery is significantly more efficient than the association rule discovery independent of data structure and implementation. Optimal rule discovery is an efficient alternative to association rule discovery, especially when the minimum support is low.