Evidential Graph Contrastive Alignment for Source-Free Blending-Target Domain Adaptation.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Domain Adaptation (DA) faces challenges with real-world data heterogeneity.
• Existing DA methods struggle with source-free settings and multiple unlabeled target domains.
• Source-free Blending-Target Domain Adaptation (SF-BTDA) presents a realistic yet complex scenario with mixed label shifts and noisy pseudo-labels.

Purpose of the Study:

• To propose a novel method, Evidential Graph Contrastive Alignment (EGCA), for the SF-BTDA setting.
• To address the challenges of inaccessible source data, unlabeled multiple target domains, and noisy pseudo-labels.
• To improve the accuracy and certainty of pseudo-target labels and minimize distribution gaps within blended target domains.

Main Methods:

• Calibrated Evidential Learning (CEL) module for iterative improvement of model accuracy and certainty, generating high-quality pseudo-target labels.
• Graph contrastive learning incorporating a domain distance matrix and confidence-uncertainty criterion to align sample distributions across blended targets.
• Development of a new benchmark using three standard DA datasets to evaluate the proposed method.

Main Results:

• EGCA significantly outperforms existing methods in the SF-BTDA setting.
• The proposed method achieves considerable performance gains.
• EGCA demonstrates comparable results to methods with access to source data or domain labels.

Conclusions:

• EGCA offers an effective solution for the challenging SF-BTDA problem.
• The method successfully decouples the blending-target domain and mitigates the impact of noisy pseudo-labels.
• EGCA advances the state-of-the-art in domain adaptation for complex, real-world scenarios.