Leveraging Group-Level Signals for Robust Many-Domain Generalization

Bohan Zhang, Yachuan Liu, Qiaozhu Mei, Paramveer S. Dhillon

Abstract: Many-domain generalization poses a significant challenge for machine learning models, as real-world data often comprises numerous domains in both training and unseen test sets. Traditional approaches that focus on a limited number of domains fail to capture this complexity, leading to suboptimal performance. We propose a novel group-wise reweighting strategy that leverages diverse group-level features—including label entropy, representation statistics, and gradient properties—to address this problem. By determining group importance during training based on these features, our approach overcomes limitations of existing methods that rely solely on group error. Our results demonstrate significant improvements in worst-group and tail performance on out-of-sample test data across multiple datasets, validating our selective upweighting strategy. We further implement a learning-to-rank framework that integrates multiple group features. While this approach achieves substantial gains over empirical risk minimization, the challenges in consistently outperforming individual features highlight the inherent difficulties in achieving transferable robustness amid varying group characteristics across datasets. Finally, SHAP analysis confirms the heterogeneous importance of different features, emphasizing the need for adaptive strategies.

External IDs:dblp:conf/pakdd/ZhangLMD25