A Many-objective Feature Selection Algorithm for Multi-label Classification Based on Computational Complexity of Features

摘要

Multi-label classification constructs a model on instances which are associated to a set of labels. Similar to traditional single-label classification, redundant and irrelevant features degrade the performance of classification in terms of multiple criteria. So, feature selection task can be modeled as optimizing several conflicting objectives in a large search space simultaneously. Minimizing the number of features and the error of classification are two well-known objectives which are considered in several multi-objective feature selection methods. In addition of these objectives, the computational complexity of features is one of the most important properties of selected features which should be minimized as a crucial objective. As a result, it is desired to select less complex features while offer a higher classification accuracy. The key contribution of this paper is proposing a many-objective optimization method, for first time, to select best subset of features for multi-label data based on four objectives including number of features, two classification error measures (i.e., Hamming loss and Ranking loss), and the complexity of selected features. The defined many-objective optimization problem is solved using a proposed binary version of NSGA-III algorithm. In order to evaluate the proposed algorithm (i.e., binary NSGA-III), a benchmarking is conducted on eight multi-label datasets in terms of several multi-objective assessment. Experimental results show significant improvements for proposed method in comparison with NSGA-II approach.