Distribution-sensitive learning for imbalanced datasets

Abstract

Many real-world face and gesture datasets are by nature imbalanced across classes. Conventional statistical learning models (e.g., SVM, HMM, CRY), however, are sensitive to imbalanced datasets. In this paper we show how an imbalanced dataset affects the performance of a standard learning algorithm, and propose a distribution-sensitive prior to deal with the imbalanced data problem. This prior analyzes the training dataset before learning a model, and puts more weight on the samples from underrepresented classes, allowing all samples in the dataset to have a balanced impact in the learning process. We report on two empirical studies regarding learning with imbalanced data, using two publicly available recent gesture datasets, the Microsoft Research Cambridge-12 (MSRC-12) and NATOPS aircraft handling signals datasets. Experimental results show that learning from balanced data is important, and that the distribution-sensitive prior improves performance with imbalanced datasets.