Identifying Predictive Structures in Relational Data Using Multiple Instance Learning

Abstract

This paper introduces an approach for identifying predictive structures in relational data using the multiple-instance framework. By a predictive structure, we mean a structure that can explain a given labeling of the data and can predict labels of unseen data. Multiple-instance learning has previously only been applied to flat, or propositional, data and we present a modification to the framework that allows multiple-instance techniques to be used on relational data. We present experimental results using a relational modification of the diverse density method and of a method based on the chi-squared statistic. We demonstrate that multipleinstance learning can be used to identify predictive structures on both a small illustrative data set and the Internet Movie Database. We compare the classification results to a k-nearest neighbor approach. ICML Proceedings of the Twentieth International Conference on Machine Learning