Nonconvex and discriminative transfer subspace learning

Unsupervised transfer subspace learning is one of the challenging and important topics in domain adaptation, which aims to classify unlabeled target data by using source domain information. The previous research still have shortcomings in constructing two domain similarity structures and avoiding the negative impact of label noise.

To deal with this problem, a research team led by Tingjin Luo published their new research on 15 Feb 2025 in Frontiers of Computer Science co-published by Higher Education Press and Springer Nature.

The team proposed a novel nonconvex and discriminative transfer subspace learning (NDTSL) to transfer knowledge from source domain to target domain and extract the discriminative features in subspace by incorporating Schatten-p norm and a soft label matrix. Except for the computational complexity analyses, the experimental results illustrate the superiority of the proposed approach.

In the research, the Schatten-p norm and soft source label matrix are adopted to tightly approximate data similarity structure and enhance data discriminability respectively, thereby effectively improving the model classification performance on target data. Specifically, to maintain the low-rank property of the similarity structure between source and target domain data, a nonconvex Schatten-p norm is used to approximate the rank function rather than the trace norm, obtaining a better low-rank representation. Meanwhile, the transfer subspace learning model is sensitive to label noise when fitting source labels. A soft label matrix is adopted to learn a more flexible classifier, which not only avoids the negative impact of label noise but also enhances the discriminability of target data. Besides, an alternative inexact augmented Lagrange multiplier method is designed to solve the nonconvex objective function. Extensive experiments on eighteen transfer tasks, two classifiers, and four evaluation metrics clearly validate the effectiveness of the NDTSL method.

In the future, the generation mechanism of label noise will be considered for designing a robust model. Furthermore, methods to align the subspace and distribution of domains simultaneously and extend the method to solve the problem of incomplete features are worth studying.

DOI: 10.1007/s11704-023-3228-0