Open continual sampling with hypersphere knowledge transfer for rapid feature selection

Abstract

Feature selection is a widely used data preprocessing technique, but it still faces two major challenges: (1) data in open and dynamic environments may continually emerge unknown classes, and (2) the ever-growing scale of data. To address these challenges, this paper proposes a novel Open Continual Sampling (OCS) method that combines the advantages of continual learning and three-way sampling, aiming to discover unknown knowledge and transfer known knowledge. OCS can detect unknown classes by constructing a hypersphere knowledge base and sampling the most uncertain instances at each class decision boundary from the unknown data, thereby effectively reducing redundant sample computations. Based on OCS, we introduce a rapid feature selection framework (OCS-FS). Guided by the prior knowledge base, this framework rapidly calculates the importance of a small number of candidate features on representative samples, thereby incrementally selecting the optimal feature subset for the new data. After completing the learning process for the new period, the knowledge base is updated to reinforce old knowledge and integrate new knowledge. Extensive experiments on public benchmark datasets demonstrate that our method significantly outperforms existing state-of-the-art feature selection methods in both effectiveness and efficiency.