Simultaneous Latent Budget Trees offer explainable AI for stratified classification

Researchers from the University of Naples Federico II have published a new approach to decision tree classification that tackles a persistent problem: how to handle stratification factors like time, location, or demographic variables that act as confounders. Their Simultaneous Latent Budget Trees (SLBT) framework replaces standard impurity-based splits with a model-based rule that treats child nodes as latent components of a Simultaneous Latent Budget Model. This allows the tree to optimize for conditional splits that separate the effect of the stratification variable from other predictors, while maintaining the interpretability that makes single trees attractive for explainable AI.

The technical innovation comes from recasting the tree growth procedure as a neural network-like optimization problem, estimating mixing parameters and latent budget profiles via least squares. The resulting tree structure can be interactively visualized with aids for pruning and path analysis. The method also includes measures to handle unbalanced response classes. The authors demonstrate SLBT on a real-world dataset of Amyotrophic Lateral Sclerosis (ALS) progression, successfully identifying gender-related differences in disease trajectories. An accompanying open-source library (SLBT) is available on GitHub, making the algorithms accessible for other researchers and practitioners working with stratified classification problems.