Evolutionary Feature Selection based on Curriculum Learning for Large-instance Datasets

Among wrapper-based feature selection, evolutionary computation is considered as one of the most promising solutions and has became increasingly popular. However, it faces a serious scalability challenge, especially when dealing with large-instance datasets that are very widespread in the era of big data. To address this issue, this paper proposes a feature selection algorithm tailored for large-instance datasets, namely, curriculum learning based evolutionary feature selection (called CLefs for short). As far as we know, it’s the first work combining curriculum learning with evolutionary computation. By using a model-agnostic subset selection method and an adaptive mechanism for adjusting the subset size, a curriculum is constructed based on a series of easy-to-hard subsets for fitness evaluations, greatly reducing the heavy computational cost of evolutionary feature selection algorithms without sacrificing accuracy. A series of experiments are conducted on several large datasets to verify the effectiveness and efficiency of the proposed CLefs algorithm. Empirical results demonstrate that our proposed CLefs algorithm significantly outperforms seven peer algorithms in most cases, and can efficiently tackle feature selection on large-instance datasets.