Semi-supervised minimum redundancy maximum relevance feature selection for audio classification

Abstract

It is still a changing problem of choosing the most relevant ones from multiple features for their specific machine learning tasks. However, feature selection provides an effective solution to it, which aims to choose the most relevant and least redundant features for data analysis. In this paper, we present a feature selection algorithm termed as semi-supervised minimum redundancy maximum relevance. The relevance is measured by a semi-supervised filter score named constraint compensated Laplacian score, which takes advantage of the local geometrical structures of unlabeled data and constraint information from labeled data. The redundancy is measured by a semi-supervised Gaussian mixture model-based Bhattacharyya distance. The optimal feature subset is selected by maximizing feature relevance and minimizing feature redundancy simultaneously. We apply our algorithm in audio classification task and compare it with other known feature selection methods. Experimental results further prove that our algorithm can lead to promising improvements.