Model-Agnostic Meta Learning for Class Imbalance Adaptation

A team of researchers has introduced HAMR (Hardness-Aware Meta-Resample), a novel framework designed to solve the pervasive problem of class imbalance in natural language processing (NLP). Class imbalance, where some categories have far fewer training examples than others, severely degrades model performance on critical minority classes in domains like medical diagnosis or disaster response. HAMR employs a two-pronged, model-agnostic approach using meta-learning principles. It performs bi-level optimization to dynamically estimate instance-level weights, prioritizing both genuinely difficult samples and underrepresented classes. Simultaneously, a neighborhood-aware resampling mechanism amplifies the training signal for these hard examples and their semantically similar neighbors, ensuring the model learns robust features rather than just memorizing rare cases.

The framework was rigorously tested on six imbalanced datasets covering multiple NLP tasks across biomedical, disaster response, and sentiment analysis domains. Experimental results demonstrate that HAMR achieves substantial improvements for minority classes, consistently outperforming strong existing baselines. Extensive ablation studies confirmed that both of its core modules—the dynamic weighting and the resampling mechanism—synergistically contribute to these performance gains. The research, accepted to the Findings of ACL 2026, provides a unified and generalizable tool for developers. By making the code publicly available, the team offers a practical solution for building more equitable and reliable AI systems that perform well not just on common data, but on the rare, often critical, cases that matter most.