Resampling and Cost-Sensitive Methods for Imbalanced Multi-instance Learning
Xiaoguang Wang, Xuan Liu, Nathalie Japkowicz, Stan Matwin · 2013
Multi-instance learning uses a set of bags containing many instances, which makes it different from standard propositional classification. Our research shows that, similar to the single-instance imbalance problem, classification of multi-instance data with imbalanced class distributions significantly degrades performance when compared to most standard multi-instance algorithms in a balanced setting. Due to the inherent differences between multi-instance and single-instance learning, the existing solutions for single-instance class imbalance problems do not transfer directly to multi-instance datasets. This is a drawback, as imbalanced multi-instance problems often occur in data mining practice. In this paper, we propose two solution frameworks for multi-instance class imbalanced datasets. In the first we explore multi-instance data sampling methods, and in the second we present a novel generalized version of a multi-instance cost-sensitive boosting technique. Experimental results, on benchmark datasets and application datasets, show that the proposed frameworks are an effective solution for the multi-instance class imbalance problem.