Track papers across labs, topics, and trends.
Search papers, labs, and topics across Lattice.
This paper introduces neural classification trees (NCT), a novel framework designed to enhance classification robustness by encoding latent subgroup structures within a tree architecture. By dynamically routing samples through "easy" or "hard" nodes based on prediction correctness and reusing these routes as pseudo-labels, NCT effectively disentangles conflicting subgroups without requiring explicit subgroup supervision. Evaluations across five benchmarks reveal that NCT not only improves interpretability by isolating minority subgroups but also achieves competitive robustness compared to existing state-of-the-art methods.
NCT offers a groundbreaking approach to classification that reveals hidden subgroup structures while maintaining high robustness, all without needing subgroup annotations.
Machine learning models exploit spurious correlations, achieving high average accuracy but failing disproportionately on underrepresented subgroups. Existing methods address this by adjusting network parameters, guided either by subgroup annotations or inferred pseudo-group labels. Yet at inference, these methods produce only a class prediction, with no insight into a sample's latent subgroup. We propose neural classification trees (NCT), a framework that achieves robustness by encoding subgroup structure in its tree-shaped architecture. By routing each sample to an "easy" or "hard" node of this tree -- based on prediction correctness -- and reusing these routes as pseudo-labels for the next iteration, NCT disentangles conflicting subgroups, without requiring subgroup supervision. We evaluate NCT on five benchmarks spanning binary and multi-class spurious correlations. Our experiments show that the learned tree topology provides strong interpretability by consistently isolating minority subgroups, which provides a transparent mapping between the model architecture and the data's latent group structure, while yielding competitive robustness with state-of-the-art methods.