Building predictive models for tabular data presents fundamental challenges, notably in scaling consistently, i.e., more resources translating to better performance, and generalizing systematically beyond the training data distribution. Designing decision tree models remains especially challenging given the intractably large search space, and most existing methods rely on greedy heuristics, while deep learning inductive biases expect a temporal or spatial structure not naturally present in tabular data. We propose a hybrid amortized structure inference approach to learn predictive decision tree ensembles given data, formulating decision tree construction as a sequential planning problem. We train a deep reinforcement learning (GFlowNet) policy to solve this problem, yielding a generative model that samples decision trees from the Bayesian posterior. We show that our approach, DT-GFN, outperforms state-of-the-art decision tree and deep learning methods on standard classification benchmarks derived from real-world data, robustness to distribution shifts, and anomaly detection, all while yielding interpretable models with shorter description lengths. Samples from the trained DT-GFN model can be ensembled to construct a random forest, and we further show that the performance of scales consistently in ensemble size, yielding ensembles of predictors that continue to generalize systematically.

The standard approach of learning decision trees from histogram data is to treat the bins as independent variables. However, as the underlying dependencies among the bins might not be completely exploited by this approach, an algorithm has been proposed for learning decision trees from histogram data by considering all bins simultaneously while partitioning examples at each node of the tree. Although the algorithm has been demonstrated to improve predictive performance, its computational complexity has turned out to be a major bottleneck, in particular for histograms with a large number of bins. In this paper, we propose instead a sliding window approach to select subsets of the bins to be considered simultaneously while partitioning examples. This significantly reduces the number of possible splits to consider, allowing for substantially larger histograms to be handled. We also propose to evaluate the original bins independently, in addition to evaluating the subsets of bins when performing splits. This ensures that the information obtained by treating bins simultaneously is an additional gain compared to what is considered by the standard approach. Results of experiments on applying the new algorithm to both synthetic and real world datasets demonstrate positive results in terms of predictive performance without excessive computational cost.