Contrastive Semi-Supervised Domain Adaptation with von Mises-Fisher for Class Imbalance Mitigation

MENG-ZHU WANG, Yin Nan

Keywords: Semi-supervised Domain Adaptation, Distributional Discrepancy, Contrastive Learning, von Mises-Fisher

TL;DR: cSSDA

Abstract: In contrast to Unsupervised Domain Adaptation (UDA) methods that rely solely on unlabeled data, Semi-supervised Domain Adaptation (SSDA) aims to enhance classification accuracy and generalization by incorporating a small amount of labeled samples from the target domain. However, a central challenge in SSDA is effectively addressing the distributional discrepancy between the source and target domains, particularly when labeled target data is scarce. While existing SSDA approaches have alleviated this issue to some extent, the persistent problem of class imbalance remains a critical obstacle. To address this challenge, we propose a novel Contrastive Semi-Supervised Domain Adaptation (cSSDA) algorithm based on the von Mises-Fisher (vMF) distribution. The core idea is to integrate the vMF distribution into the contrastive learning framework to refine the contrastive loss, enabling the construction of an infinite number of contrastive pairs. This approach helps the model better handle the class imbalance inherent in SSDA. Specifically, the vMF distribution excels at modeling directional data in high-dimensional spaces, enhancing the model's ability to capture similarities and differences between source and target domains during contrastive learning. Extensive experiments conducted on two widely-used benchmark datasets demonstrate that our method consistently outperforms existing SSDA approaches.

Primary Area: transfer learning, meta learning, and lifelong learning

Submission Number: 1655