A deep reinforcement learning-based approach for dynamic feature selection in multi-label streaming data (SMLFS-DRL)

Multi-label classification techniques assign multiple labels to each instance and are widely applied in real-world tasks such as cancer diagnosis and text categorization. Traditional methods face significant limitations when dealing with the dynamic nature of streaming data. Although feature selection techniques have been developed to address this issue, conventional approaches that assume prior access to all features remain ineffective in streaming environments. This study proposes a Deep Reinforcement Learning (DRL)-based unified framework for dynamic feature selection in multi-label streaming data. In the proposed approach, DRL serves as the core decision-making mechanism that sequentially selects features in an online manner. Mutual information and Graph Neural Networks (GNNs) are integrated within the DRL environment to model feature–label relevance and feature interactions, respectively, forming an informative state representation and reward structure. Furthermore, the NSGA-II algorithm is employed as a multi-objective optimization component to dynamically update the Pareto front of candidate feature subsets. The key objectives are to maximize feature relevance to the labels, minimize feature redundancy, and maintain an adaptive Pareto front to ensure the selection of optimal feature subsets over time. Experimental results on eleven benchmark datasets demonstrate that the proposed method outperforms state-of-the-art Online Feature Selection (OLFS) techniques in terms of predictive accuracy, statistical significance, and stability.

کلیدواژه ها:

Streaming Data Feature selection ، Online Multi-Label Learning ، Deep reinforcement learning ، Neural Networks ، Multi-Objective Optimization ، Mutual Information