Learning Rules from Multiple Instance Data: Issues and Algorithms

In a multiple-instance representation, each learning example is represented by a “bag” of fixed-length “feature vectors”. Such a representation, lying somewhere between propositional and first-order representation, offers a tradeoff between the two. This paper proposes a generic extension to propositional rule learners to handle multiple-instance data. It describes NAIVE-RIPPERMI, an implementation of this extension on the rule learning algorithm RIPPER. It then explains several pitfalls encountered by this naive extension during induction. It goes on to describe algorithmic modifications and a new multipleinstance coverage measure which are shown to avoid these pitfalls. Experimental results show the benefits of this approach for solving propositionalized relational problems in terms of speed and accuracy. keywords: Multiple-instance learning problem, rule learning, propositionalization, relational learning, mutagenesis learning task