Gradual spatial constraint feature selection for robust biomarker discovery in high-dimensional gene expression data

High-dimensional microarray gene expression data presents unique challenges for feature selection to identify sparse, biologically relevant biomarkers conflicts with computational constraints. To overcome these limitations, this paper proposes a Gradual Spatial Constraint Feature Selection framework (GSCFS), a novel wrapper method specifically designed for genomic data that strategically balances search space compression with functional feature interaction modeling. The framework introduces three key innovations. First, a two-stage gradual spatial constraint mechanism based on dimension prioritization, which dynamically controls the decision space of candidate solutions to improve optimization efficiency. Second, a spatial convergence operator employing a chain-based convergence strategy, enabling self-adaptive individual updates based on historical states—reducing external dependencies while accelerating convergence and computational efficiency. Finally, spatial diffusion operator integrating three synergistic strategies (group-space constraints, elite-space constraints, and spatial-domain constraints) to facilitate hierarchical feature co-optimization and enhance feature interactions. Extensive experiments on real-world high-dimensional gene expression datasets demonstrate the statistically validated superiority of the proposed approach. Statistical analysis confirms that GSCFS ranks first in classification accuracy among nine state-of-the-art methods. It successfully isolates minimal feature subsets without compromising predictive performance. Furthermore, the framework achieves an order-of-magnitude reduction in computational time compared to traditional evolutionary wrappers, effectively resolving the conflict between biomarker sparsity and search efficiency.