A boosting approach to multiple instance learning

Abstract

In this paper we present a boosting approach to multiple instance learning. As weak hypotheses we use balls (with respect to various metrics) centered at instances of positive bags. For the co-norm these hypotheses can be modified into hyper-rectangles by a greedy algorithm. Our approach includes a stopping criterion for the algorithm based on estimates for the generalization error. These estimates can also be used to choose a preferable metric and data normalization. Compared to other approaches our algorithm delivers improved or at least competitive results on several multiple instance benchmark data sets. © Springer-Verlag Berlin Heidelberg 2004.