Does Source Data Harm Adaptation? A Mixed-Supervised Degradation-Aware Framework

This paper studies two crucial undiscovered issues for semi-supervised domain adaptation (SSDA) that severely limit existing methods. Firstly, their development has been confined to the transductive semi-supervised learning paradigm, neglecting the potential of other paradigms to improve adaptation. Secondly, contrary to common assumption, increasing source domain training data often degrades performance, a phenomenon we term Source-Induced Degradation (SID). To address these issues, we propose a novel dictionary pair transfer learning method under a mixed-supervised learning paradigm (DPTL-Mixed). DPTL-Mixed is built upon the discriminative Dictionary Pair Learning (DPL) framework, leveraging its linear efficiency and representational power. Our framework consists of two complementary components: a supervised component (DPTL-Sup) and a semi-supervised component (DPTL-Semi). DPTL-Sup learns separate domain-specific analysis dictionaries for source and target data, enabling explicit re-weighting of domain contributions to mitigate SID. DPTL-Semi follows the traditional SSDA philosophy, incorporating distribution alignment via adaptive maximum mean discrepancy and manifold regularization. The final model fuses the outputs of both components, levering the complementary strengths of supervised and semi-supervised learning. Experimental results show that the proposed method outperforms other shallow competitors in 53 out of 60 SSDA tasks, validating its effectiveness in overcoming both identified issues.