Adaptive High-Dimensional Subspace Evolution Based on Broad Learning System and Error-Correcting Output Codes

Gongli Zhang, zhulin liu, C.L.Philip Chen, Tong Zhang, Guang-Ze Chen

Keywords: high-dimensional data, subspace evolution, broad learning system, error-correcting output codes

Abstract: High-dimensional data (HDD) commonly exhibit complex hierarchical structural characteristics; however, existing approaches typically employ fixed subspace evolution strategies that fail to adapt to the inherent hierarchical diversity across different datasets, resulting in suboptimal revelation of underlying discriminative patterns. Considering this critical limitation, we propose an adaptive high-dimensional subspace evolution algorithm (AHSE) featuring a dual-branch collaborative architecture: the series branch leverages Cholesky decomposition-based incremental Broad Learning System (BLS) to efficiently evolve cascaded subspaces tailored to distinct types of high-dimensional hierarchies; the parallel branch, built on multiple subspace evolution bases, utilizes post-hoc error-correcting output codes (ECOCs) for robust spatial encoding and evolutionary optimization. Both branches converge into a lightweight circuit, forming a closed evolutionary loop. Owing to the hierarchy-tailored evolution strategy, AHSE excels in various HDD tasks such as image pattern recognition, speech emotion recognition, and few-shot learning. Moreover, we offer a rigorous theoretical analysis of the mechanism and robustness guarantee of ECOCs on BLS, further promoting the integrity of AHSE.

Primary Area: optimization

Submission Number: 23588