RuleExplorer: A Scalable Matrix Visualization for Understanding Tree Ensemble Classifiers.

Area of Science:

• Computer Science
• Artificial Intelligence
• Data Visualization

Background:

• Tree ensemble classifiers achieve high performance using numerous rules, but this complexity hinders interpretability.
• Existing model reduction techniques simplify classifiers by extracting rule subsets, often losing crucial information and ignoring infrequent but important anomalous rules.

Purpose of the Study:

• To develop a scalable visual analysis method for explaining tree ensemble classifiers with tens of thousands of rules.
• To enhance model interpretability by preserving fidelity and incorporating anomalous rules.

Main Methods:

• Adaptive hierarchical organization of rules to maintain comprehensiveness.
• Anomaly-biased model reduction to prioritize infrequent but critical rules at each level.
• Matrix-based hierarchical visualization for multi-level exploration of rules.

Main Results:

• The proposed method effectively explains tree ensemble classifiers by organizing rules hierarchically.
• It successfully incorporates and highlights anomalous rules, which are often missed by traditional methods.
• Quantitative experiments and case studies validate the method's ability to foster a deeper understanding of classifier logic.

Conclusions:

• The developed visual analysis method enhances the interpretability of complex tree ensemble models.
• It achieves this by preserving model fidelity through hierarchical organization and prioritizing anomalous rules.
• This approach offers a comprehensive understanding of both common and anomalous decision pathways within the classifier.