Imbalanced Classification through the Lens of Spurious Correlations

Imbalanced datasets can cause AI models to rely on misleading patterns: The paper reveals that when training data has unequal class sizes, models tend to learn "Clever Hans effects"—spurious correlations that appear to work but don't reflect true causal relationships. For example, a model might classify images based on background details rather than the actual subject, leading to failures in real-world deployment.

Traditional reweighting approaches miss the root problem: While existing methods try to fix class imbalance through data or loss reweighting, this paper argues these approaches don't address the underlying issue—that minority classes are underspecified and prone to spurious pattern-learning. The authors propose a fundamentally different perspective on why imbalance causes problems.

Explainable AI can help identify and fix unreliable patterns: The paper demonstrates that using counterfactual explanations (a type of interpretability tool) can help practitioners jointly identify which spurious correlations their models have learned and remove them, rather than just accepting poor performance as inevitable.

The problem is more widespread than previously recognized: The authors note that the connection between class imbalance and spurious correlations has been "largely overlooked by existing approaches," suggesting this is an underappreciated source of AI reliability problems in real-world applications with imbalanced data.

Competitive performance is achievable while addressing the root cause: The method achieved competitive classification results on three datasets while explicitly tackling spurious correlations, suggesting organizations don't have to choose between accuracy and reliability.