Anti-drifting Feature Selection via Deep Reinforcement Learning (Student Abstract)

Feature selection (FS) is a crucial procedure in machine learning pipelines for its significant benefits removing data redundancy and mitigating model overfitting. Since concept drift widespread phenomenon streaming could severely affect performance, effective FS on drifting streams imminent. However, existing state-of-the-art algorithms fail to adjust their strategy adaptively when the feature subset changes, making them unsuitable streams. In this paper, we propose dynamic method that selects features via deep reinforcement learning. Specifically, present two novel designs: (i) skip-mode environment shrinks action space size high-dimensional tasks; (ii) curiosity mechanism generates intrinsic rewards address long-horizon exploration problem. The experiment results show our proposed outperforms other methods can dynamically adapt drifts.