Multi-objective PSO based online feature selection for multi-label classification

Abstract

Feature selection approaches aim to select a set of prominent features that best describe the data to improve the efficiency without degrading the performance of the model. In many real-world applications such as social networks, it is not easy to get a static feature set; rather, new features arrive continuously in the system. Therefore, online feature selection (OFS) strategies have become popular in dealing with such problems. Recent years have also witnessed the prominence of multi-label classification frameworks where multiple class labels can be associated with a single instance. The proposed method considers the multi-label learning and the arrival of features in an online fashion. The method automatically determines the best subset of features that is suitable for multi-label classification. A three-phase filtering process is applied to select the appropriate features. The first phase is an evolutionary-based particle swarm optimization (PSO) technique that applies to the group of incoming features in a multi-objective framework. The second phase checks the redundancy of features selected in the current group to the already selected features and finally, the third phase finds the features in the already selected feature list that becomes non-significant on the selection of newly arrived features and discards them. The proposed algorithm is tested on fourteen multi-label data sets collected from various domains such as biology, music, and text. From the results, it is observed that the first and second phases are sufficient to select the appropriate feature set. The efficacy of the proposed algorithm can be verified from the obtained results. It outperforms the results obtained by state-of-the-art approaches in most cases.