S³DL: Sample-Aggregated Structured Supervised Dictionary Learning

This paper proposes Sample-aggregated Structured Supervised Dictionary Learning (S³DL), a novel framework for robust classification in noisy and outlier-prone scenarios. S³DL jointly optimizes synthesis and analysis dictionaries, along with a meta-sample aggregation projection and adaptive weighting matrices, to learn a discriminative latent feature space that suppresses noise and outliers while enhancing classification. The resulting constrained multi-objective optimization problem is efficiently solved via a tailored alternating optimization algorithm. Furthermore, an error bound is derived to guarantee its robustness and efficiency. Extensive experiments on seven benchmarks demonstrate that S³DL consistently outperforms state-of-the-art dictionary learning methods in classification accuracy.