Making Tree Ensembles Interpretable: A Bayesian Model Selection Approach

Satoshi Hara, Kohei HayashiAISTATS 2018 · OpenReview

Abstract

Tree ensembles, such as random forests, are renowned for their high prediction performance. However, their interpretability is critically limited due to the enormous complexity. In this study, we propose a Bayesian model selection approach to simplify these complex ensembles into a handful of human-readable rules without sacrificing significant predictive performance.

Introduction

Tree ensembles like Random Forests and XGBoost are the gold standard for tabular data because they divide the input space into thousands of tiny, precise regions. However, this high accuracy comes at the cost of interpretability; a model with 1,000+ rules is essentially a "black box" that is impossible for a human to audit.

The core issue is fragmentation. Even when the underlying data has a simple structure, an ensemble of many trees creates a "shattered" prediction surface with massive redundancy. This paper introduces a Bayesian "defragging" method to consolidate these thousands of tiny regions into a few large, representative regions that capture the essence of the model.

Visualizing the defragmentation process: from complex tree ensembles to simplified models.

Methods

The authors treat the simplification of a tree ensemble as a Bayesian model selection problem. Their approach involves three key steps:

  1. Probabilistic Representation: The deterministic tree ensemble is transformed into a probabilistic generative model. The input space is represented using a binary feature vector derived from all split points in the original trees.
  2. FAB Inference: To find the simplest model that fits the data, the researchers use Factorized Asymptotic Bayesian (FAB) inference. Unlike standard algorithms that require a pre-defined number of rules ($K$), FAB inference applies the "Occam’s razor" principle to automatically prune redundant regions during learning.
  3. Scalability via Sampling: To handle ensembles with tens of thousands of split points, a "sampling FAB" heuristic is used. By randomly sampling a subset of split points, the method effectively removes redundancy (e.g., $x < 0.5$ vs. $x < 0.51$) without losing significant information.

Results

The method was evaluated on synthetic datasets and real-world benchmarks including Spambase, MiniBooNE, and Higgs.

Comparison of FAB and EM algorithms across datasets, showing test error relative to the number of rules (K).

Key findings include:

  • Interpretability: On the Energy dataset, a complex ensemble was reduced to just four rules while maintaining high accuracy.
  • Efficiency: The FAB inference approach was 3 to 7 times faster than the standard EM algorithm because it avoids manual searching over different values of $K$.
  • Performance: The method produced significantly fewer rules (usually <10) than other rule-extraction methods like inTrees, while achieving better accuracy than simple decision trees.

Discussion and Conclusion

The success of this approach highlights that tree ensembles are incredibly redundant; one can discard 99% of split points and still reconstruct an interpretable version of the model. By using Bayesian marginal likelihood, the algorithm naturally favors simpler models, automatically truncating rules that do not contribute significantly to the model's evidence.

While the sampling heuristic is effective, it remains a heuristic with a theoretical risk of missing rare but crucial split points. Future work may extend this Bayesian framework to other black-box architectures beyond tree ensembles.

Summary updated

Abstract

Tree ensembles, such as random forests, are renowned for their high prediction performance. However, their interpretability is critically limited due to the enormous complexity. In this study, we p...
BibTeX
@inproceedings{hara2018making,
  author = {Satoshi Hara and Kohei Hayashi},
  title = {{Making Tree Ensembles Interpretable: A Bayesian Model Selection Approach}},
  booktitle = {International Conference on Artificial Intelligence and Statistics},
  year = {2018},
  url = {https://openreview.net/forum?id=AYwsSvT44hO},
}
Figure