Maximum Relevant Minimum Redundant Multi-Label Feature Selection Using Ant Colony Optimization

Abstract

Multi-label classifiers assign more than one label to each instance, but their performance is reduced when faced with high-dimensional tasks. This work introduces a multi-layered graph modeling framework that incorporates ant colony optimization for efficient multi-label feature selection, named Maximum Relevant Minimum Redundant Multi-Label Feature Selection (MR2MLFS). The proposed method first maps the feature space into a multi-layer graph. In the first layer, a graph clustering method is used to group correlated features. Then, a meta-graph is generated to represent the relationship between clusters in the second layer. We modify the search process of ant colony optimization so that it moves between nodes in the first layer with high probability while searching within similar nodes in the second layer with low probability. In this way, the ants select highly relevant and non-redundant features. Using a multi-label classifier in the search process typically incurs high computational costs to reach the final solutions. To address this issue, we propose an information-theoretic measure to assess the identified feature sets. This measure computes both the correlation of features with target labels and their internal correlation. We evaluated the proposed method from various aspects, including efficiency, statistical performance, and stability, through tables and charts, all of which demonstrate the superiority of the proposed method compared to other approaches. Also, we evaluate the effectiveness of the proposed method on a set of real-world applications, and the results show its superiority over state-of-the-art methods.

Keywords: Feature selection, Multi-Label Data, Mutual Information, Graph Clustering, Ant Colony Optimization