Multi-Branch Neural Decision Trees Improve Shallow Tree Accuracy.

Decision trees are valued for their interpretability in tabular prediction tasks, but conventional binary trees can lack expressiveness when constrained by a strict maximum depth. Researchers have introduced the Multi-Branch Neural Decision Tree with Adaptive Pruning (MBNDT) to address this limitation, aiming to improve accuracy while maintaining short root-to-leaf paths. MBNDT is an axis-aligned tree trained end-to-end using differentiable multi-way splits. Each internal node learns multiple ordered thresholds over a selected feature and incorporates a branch mask that dynamically adjusts its effective arity. After training, the model converts to a deterministic single-path tree for inference. Experiments across 21 binary-classification benchmarks show MBNDT achieving superior average rank and mean balanced accuracy compared to other depth-constrained single-tree baselines, with multi-way splitting identified as the primary source of these gains. This approach is particularly suited for scenarios where accuracy under short, bounded decision paths is prioritized over minimal global tree size.