Transfer Learning: Kernel-Based Domain Adaptation with Distance-Based Penalization

Abstract

This paper introduces a novel approach to address the challenges of transfer learning, which aims to efficiently train a classifier for a new domain using supervised information from similar domains. Traditional transfer learning methods may fail to maintain the discriminative features of the target domain due to the scarcity of labelled data and the use of irrelevant source domain data distribution subspace, resulting in poor metrics. To overcome these challenges, the proposed approach, called KDADP, transforms the data distribution of both the source and target domains into a lower-dimensional subspace while preserving their discriminatory information. The KDADP model maximizes between-class variance and minimizes within-class variance with L1 penalization, enabling the recovery of the most useful characteristics and reducing the model’s complexity. Experimental results on three real-world domain adaptation datasets demonstrate that the proposed KDADP model significantly improves classification performance and outperforms state-of-the-art primitive, shallow, and deeper domain adaptation methods.