RuleExplorer: A Scalable Matrix Visualization for Understanding Tree Ensemble Classifiers

TL;DR

A scalable visual analysis method to explain tree ensemble classifiers that contain tens of thousands of rules by adaptively organizing the rules as a hierarchy rather than reducing them and develops an anomaly-biased model reduction method to prioritize these rules at each hierarchical level.

Abstract

The high performance of tree ensemble classifiers benefits from a large set of rules, which, in turn, makes the models hard to understand. To improve interpretability, existing methods extract a subset of rules for approximation using model reduction techniques. However, by focusing on the reduced rule set, these methods often lose fidelity and ignore anomalous rules that, despite their infrequency, play crucial roles in real-world applications. This paper introduces a scalable visual analysis method to explain tree ensemble classifiers that contain tens of thousands of rules. The key idea is to address the issue of losing fidelity by adaptively organizing the rules as a hierarchy rather than reducing them. To ensure the inclusion of anomalous rules, we develop an anomaly-biased model reduction method to prioritize these rules at each hierarchical level. Synergized with this hierarchical organization of rules, we develop a matrix-based hierarchical visualization to support exploration at different levels of detail. Our quantitative experiments and case studies demonstrate how our method fosters a deeper understanding of both common and anomalous rules, thereby enhancing interpretability without sacrificing comprehensiveness.

Figures (10)

• Figure 1: An illustration of a tree ensemble classifier used to predict stock price movement with labels "increase", "decrease", and "stable."
• Figure 2: RuleExplorer overview: Given a set of rules from a tree ensemble classifier, the anomaly-biased model reduction calculates the anomaly score for each rule and then extracts the representative rules. Next, the representative rules are fed into the matrix-based hierarchical visualization for exploration and analysis, where a rule hierarchy is dynamically built based on user selections. The brown color indicates the analysis of the representative rules at the top level, and the purple color indicates the iterative analysis subsequently.
• Figure 3: An example illustration of multi-class hinge loss.
• Figure 4: Dynamic rule hierarchy construction between consecutive levels.
• Figure 5: RuleExplorer: (a) attribute view shows the attribute distribution and enables sample filtering; (b) matrix view shows the representative rules at a certain level of the rule hierarchy; (c) info view shows the overall statistics of the displayed rules and samples; (d) data table lists the samples covered by the displayed rules.