Towards Multi-label Feature Selection by Instance and Label Selections

摘要

In multi-label learning, feature and instance selection represent two effective dimensionality reduction techniques, which remove noise, irrelevant and redundant entries from original data for easy later analysis, such as clustering and classification. Label selection also plays a fundamental role in the pre-processing step since label-noises could negatively affect the performance of the underlying learning algorithms. The literature has been mainly limited to feature and/or instance selection, but has somewhat overlooked label selection. In this paper, we introduce, for the first time, a combination of the three selection techniques (feature, instance and label) for multi-label learning. We propose an efficient convex optimization based algorithm that evaluates the usefulness of features, instances and labels in order to select the most relevant ones, simultaneously. Experimental results on some known benchmark datasets are presented to demonstrate the performance of the proposed method.