Adaptive graph-evolutionary framework for dynamic feature refinement in multi-label learning

Multi-label learning in high-dimensional environments is fundamentally challenged by complex interdependencies among features and labels, as well as redundancy and noise that degrade predictive performance and scalability. Conventional feature selection techniques typically rely on static optimization or one-shot subset identification, limiting their ability to adapt to evolving feature relevance during learning. To address these limitations, this study introduces an adaptive graph-evolutionary framework for dynamic feature refinement in multi-label learning. Instead of treating feature selection as a fixed optimization problem, the proposed approach models feature reduction as a sequential decision process guided by structural relationships and uncertainty estimation. A heterogeneous graph is constructed to capture feature-label-instance interactions, while a graph neural network dynamically evaluates feature reliability through confidence and uncertainty modeling. These signals are integrated into an evolutionary optimization mechanism that adaptively refines the feature space through staged elimination, enabling continuous adjustment to changing label dependencies. The proposed framework is evaluated on multiple benchmark datasets using standard multi-label classifiers. Experimental results demonstrate consistent improvements over state-of-the-art methods, achieving significant reductions in Hamming Loss and Ranking Loss, alongside notable gains in Average Precision and F1-score. Statistical analysis confirms the robustness and significance of the improvements across varying data distributions.