Fast sparse supervised learning framework with BLinex loss function.

Abstract

In this paper, a novel learning model, namely the L_p-norm sparse Blinex Twin Extreme Learning Machine (PBLTELM), is proposed, which is efficient and accurate for relatively large-scale data classification. The proposed framework incorporates three key innovations: the robust Blinex loss function is integrated to enhance generalization, L_p-norm (0 < p < 1) sparsity constraints are adopted to approximate L₀-norm solutions while ensuring computational tractability, and a dual-layer optimization strategy that combines iterative weight updates with the adaptive moment estimation (Adam) algorithm is developed to address the resulting non-convex and non-smooth problem. Theoretical analysis is conducted to verify convergence to local stationary points, which ensures both computational efficiency and model accuracy and properties particularly critical for relatively large-scale applications. Comprehensive empirical evaluations are performed across diverse benchmarks, including a two-dimensional artificial dataset, the CMU facial expression dataset, 12 UCI datasets, and 7 relatively large-scale libsvm datasets, to assess the performance of PBLTELM. The results indicate statistically significant improvements in classification accuracy and computational speed compared with state-of-the-art methods, confirming that PBLTELM serves as a scalable and competitive solution for relatively large-scale classification tasks.