Discovering Latent Groups for Robust Classification

The article introduces Neural Classification Trees (NCT), a novel framework designed to enhance robustness in machine learning classification by encoding subgroup structures within a tree-shaped architecture. NCT routes samples to "easy" or "hard" nodes based on prediction correctness and iteratively uses these routes as pseudo-labels, effectively disentangling spurious correlations without requiring subgroup supervision. Evaluated across five benchmarks, NCT demonstrates strong interpretability and competitive robustness compared to state-of-the-art methods, making it a significant advancement for practitioners focused on improving model performance in the presence of underrepresented subgroups.