Class and Domain Low-rank Tensor Learning for Multi-source Domain Adaptation

Abstract

Multi-source unsupervised domain adaptation (MUDA) aims to transfer knowledge from multiple labeled source domains to an unlabeled target domain. A key challenge in MUDA is to minimize the distributional discrepancy between the source and target domains. While traditional methods typically merge source domains to reduce this discrepancy, they often overlook higher-order correlations and class-discriminative relationships across domains, which weakens the generalization and classification abilities of the model. To address these challenges, we propose a novel method called Class and Domain Low-rank Tensor Learning (CDLTL), which integrates domain-level alignment and class-level alignment into a unified framework. Specifically, CDLTL leverages a projection matrix to map data from both source and target domains into a shared subspace, enabling the reconstruction of target domain samples from the source data and thereby reducing domain discrepancies. By combining tensor learning with joint sparse and weighted low-rank constraints, CDLTL achieves domain-level alignment, allowing the model to capture complex higher-order correlations across multiple domains while preserving global structures within the data. CDLTL also takes into account the geometric structure of multiple source domains and preserves local structures through manifold learning. Additionally, CDLTL achieves class-level alignment through class-based low-rank constraints, which improve intra-class compactness and inter-class separability, thus boosting the discriminative ability and robustness of the model. Extensive experiments conducted across various visual domain adaptation tasks demonstrate that the proposed method outperforms some of the existing approaches.