.png){kind=icon}
.png){kind=icon}
Return
Towards Enhancing Prototypes Driven by Graph Convolutional Networkfor Domain Adaptation
.png){kind=icon}
PRE
AI
DOI：10.1016/j.eswa.2025.130010.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
• In this work, we first show the analysis to connect the DA theory and the proposed methodology. Based on this analysis, we provide a new perspective on the mapping space of DA by dividing it into consensus, vicinal, and vulnerable spaces. • We then find that the classification accuracy on the target domain is significantly improved when we expand the consensus and vicinal spaces while reducing the vulnerable space. To achieve this goal, we propose a novel prototype-based approach called an enhanced prototypical network (EnPro). • Next, to the best of our knowledge, the proposed methodology stands as the first prototype-based approach driven by the graph convolutional network paradigm. Within this framework, the estimated class prototypes of the Multi-Layer Perceptron (MLP) classifier undergo enhancement via pseudo-labels created from the Graph Convolutional Network (GCN) classifier. • Finally, we implemented the proposed method on several UDA and SSDA benchmark datasets such as ImageCLEF-DA, Office-31, Office-Home, VisDA-C, and DomainNet. The improved performance on the target data of the proposed method is considerable compared to previous state-of-the-art methods.
Journal
IF:
7.5
Papers:
2.9W
Citations:
10.2W