Distance-based Heuristics in Inductive Logic Programming for Mining from Chemical Compound Data

摘要

We propose an approach for making FOIL better handling multiple-instance data. This learning problem arises when trying to generate hypotheses from examples in the form of positive and negative bags. Each bag contains one or more instances and a bag is labelled as positive when there is at least one positive instance, otherwise, it is labelled as negative. Several algorithms have been proposed for learning in this framework. However all of them can only handle data in the attribute-value form which has limitations in knowledge representation. Therefore, it is difficult to handle examples consisting of structures among components, such as chemical compounds data. In this paper, the Diverse Density, a measure for multiple-instance data, is applied to adapt the heuristic function in FOIL in order to improve learning accuracy in multiple-instance data. We conducts the experiments on real-world data related to chemical compound analysis in order to show the improvement.