In transformer fault diagnosis, samples located at the classification boundaries of different operating conditions are often scarce, sparsely distributed, and partially overlapping, resulting in significant lower diagnostic accuracy than that achieved on central samples. To address this issue, this paper proposes a transformer fault diagnosis method based on XGBoost enhanced by group distributionally robust optimization (GDRO). First, Gaussian mixture model (GMM) clustering is used to cluster the samples and divide them into subclasses based on the differences in gas components in the oil, distinguishing between the central and boundary region samples. Then, an XGBoost fault diagnosis model is constructed, and GDRO algorithm is adopted to reassign weights to different subclasses, overcoming the model's tendency to underemphasize boundary samples. Finally, case studies are conducted to evaluate the performance of the proposed method. Experimental results demonstrate that, compared to conventional diagnostic methods, the proposed approach achieves a diagnostic accuracy of 91.67% for samples in boundary regions and an overall diagnostic accuracy of 91.08%, significantly enhancing the robustness and comprehensiveness of the transformer fault diagnosis model with respect to boundary samples.