Geometrical preservation and correlation learning for multi-source unsupervised domain adaptation

Multi-source unsupervised domain adaptation (MUDA) aims to improve the performance of the model on the target domain by utilizing useful information from several source domains with distinct distributions. However, due to the diverse information in each domain, how to extract and transfer useful information from source domains is essential for MUDA. Most existing MUDA methods simply minimized the distribution incongruity among multiple domains, without fully considering the unique information within each domain and the relationships between different domains. In response to these challenges, we propose a novel MUDA approach named geometrical preservation correlation learning (GPCL). Specifically, GPCL integrates graph regularization and correlation learning within the nonnegative matrix factorization (NMF) structure, leveraging the inherent geometry of the data distribution to acquire discriminative features while maintaining both the local and global geometrical structures of the original data. Meanwhile, GPCL extracts the maximum correlation information from each source domain and target domain to further narrow their domain discrepancy and ensure positive knowledge transfer. Integrated experimental results across multiple benchmarks verify that GPCL performs better than several existing MUDA approaches, showcasing the efficiency of our method in MUDA. For example, on the Office-Home dataset, GPCL outperforms the SOTA by an average of 1.58%. On the ImageCLEF-DA dataset, GPCL achieves the best results across multiple sub-tasks and the average performance, outperforming the single-source SOTA by 2.3%, 2%, and 1.26%, respectively.