Joint hierarchical multi-granularity adaptive embedding discriminative learning for unsupervised domain adaptation

Abstract

Unsupervised domain adaptation (UDA) is an effective technique that aims to transfer knowledge from well-labeled source data to target data that lacks labels and has a different distribution. Most existing methods only considered domain center-wise alignment to reduce global differences across domains, resulting in a coarse alignment. In recent years, researchers further considered aligning class centers to ensure the consistency of local distributions. However, these methods utilized a solely mean vector to represent the entire class distribution, which is still coarse and cannot fully capture the distribution characteristics of intra-class data. Inspired by the “knowledge pyramid” theory, a novel UDA method termed adaptive hierarchical multi-granularity embedded learning (HMGEL) is proposed to solve this problem, which aims to minimize the distribution gap of samples across domains from the perspective of hierarchical multi-granularity. This method can reflect the distribution of samples from coarse to fine, which is helpful for better UDA. Firstly, granular envelopes are created to explore intra-class structures and complex distributional properties at a more fine-grained level. Based on the granular envelopes, domain centers and class centers are combined for cross-domain distribution alignment, allowing for the capture of sample information at hierarchical multi-granularity from coarse to fine. Then, a robust sample-to-granular envelope cross-domain local structure learning strategy is designed to improve the discrimination capability of target domain features under hierarchical multi-granularity. Extensive experiments on five benchmark datasets show that the proposed HMGEL method is effective at a significant level.