Ensemble Feature Selection for Multi-label Classification: A Rank Aggregation Method

Abstract

Multi-label classification is an important task in machine learning applications. However, today, these methods have more challenges due to high-dimensional data. Existing methods in the literature have not yet reached an acceptable performance, and as the number of labels in the dataset increases, their performance becomes weaker. Thus providing effective feature selection methods is necessary and is suitable for all data. Ensemble feature selection is also a new approach based on combining the results of several feature selection methods. Ensemble approaches have been used in many single-label applications and have provided adequate results. However, this technique is less considered in multi-label feature selection. For this purpose, in this article, we have presented a method based on an ensemble of feature selection methods. This paper presents an ensemble multi-label feature method using rank aggregation algorithms. Three filter rankers with different structures are utilized to obtain the final ranking. This rank aggregation process is treated like an election where the rankers are assumed to be the voters, and the features are the alternatives. The Weighted Borda Count (WBC) method conducts the election process. Finally, the performance and results obtained by the proposed method are compared with the nine filter-based algorithms used based on six multi-label datasets. Most of the selected datasets have a high number of class labels, and the proposed method has recorded the best classification performance in all these datasets. Also, the results show that this method is not well affected by the adverse effects of the algorithms used.