Interpretable Attribute Discretization

Data discretization, the conversion of numeric attributes into categorical bins, underpins many data-driven tasks, from visualization and causal analysis to symbolic learning. Interpretable partitions make results easier to communicate and act upon. Yet, a fundamental tension underlies this discretization process: partitions that are interpretable to humans often differ from those that maximize utility performance (e.g., causal analysis). Despite the abundance of discretization techniques, reconciling this utility-semantics trade-off remains an open challenge. In this work, we formally define the interpretable attribute discretization problem and introduce a reinforcement-learning-based framework that jointly optimizes utility and semantic fidelity. Our approach learns a partition selection policy that can navigate a vast space of candidates and accurately estimate the optimal partitions by strategically sampling only 200 candidates out of millions. We measure success by the average Hausdorff distance between the estimated and true Pareto frontier of the utility–semantics trade-off. On seven datasets across four tasks (visualization, modeling, imputation, causal analysis) and multiple semantic measures, our method achieves the lowest overall error, typically 2–4 × smaller than strong alternatives, yielding a more accurate Pareto frontier that reveals the complete utility–semantics trade-off.