AMTEA-Based Multi-task Optimisation for Multi-objective Feature Selection in Classification

Jiabin Lin, Qi Chen, Bing Xue, Mengjie Zhang

Abstract: Feature selection is important nowadays due to many real-world datasets usually having a large number of features. Evolutionary multi-objective optimisation algorithms have been successfully used for feature selection which usually has two conflicting objectives, i.e., maximising the classification accuracy and minimising the number of selected features. However, most of the existing evolutionary multi-objective feature selection algorithms tend to address feature selection tasks independently, even when these feature selection tasks are related. Multi-task optimisation, which aims to improve the performance of multiple tasks by sharing common knowledge among them, has been used in many areas. However, there is not much work on utilising multi-task optimisation for feature selection. In this work, we develop a new multi-task multi-objective feature selection algorithm. This algorithm aims to address multiple related feature selection tasks simultaneously and facilitate knowledge capturing and transferring among the related tasks. Furthermore, a method is developed for transferring knowledge between related feature selection tasks having different features. This method can avoid transferring information between the unique features of tasks by transforming the probability models of them. We compare the proposed algorithm with the single-task multi-objective feature selection algorithm on seven sets of related feature selection tasks. Experimental results show that the proposed algorithm achieves better classification performance than the single-task algorithm with the help of knowledge transferring among related feature selection tasks. Further analysis shows that the features selected by our proposed algorithm can be more relevant to the classification tasks.