Fuzzy rough classifiers for class imbalanced multi-instance data

In multi-instance learning, each learning object consists of many descriptive instances. In the corresponding classification problems, each training object is labeled, but its constituent instances are not. The classification objective is to predict the class label of unseen objects. As in traditional single-instance classification, when the class sizes of multi-instance data are imbalanced, classification is degraded. Many multi-instance classifiers have been proposed, but few take into account the possibility of class imbalance, which causes them to fail in this situation. In this paper, we propose a new type of classifier that embodies a solution to the multi-instance class imbalance problem. Our proposal relies on the use of fuzzy rough set theory. We present two families of classifiers respectively based on information extracted at bag-level and at instance-level. We experimentally show that our algorithms outperform state-of-the-art solutions to multi-instance imbalanced data classification, evaluated by the popular metrics AUC and geometric mean.