Cluster-Based Random Forest Visualization and Interpretation.

Area of Science:

• Machine Learning
• Data Visualization
• Artificial Intelligence

Background:

• Random forests are powerful machine learning models for classification but suffer from low interpretability due to their complex structure.
• Interpreting individual decision trees within a random forest is often infeasible for understanding overall model behavior.

Purpose of the Study:

• To develop a novel visualization system and method for enhancing the interpretability of random forests.
• To enable users to grasp the general performance of a random forest model without examining every decision tree.

Main Methods:

• Introduced a new distance metric for clustering decision trees, considering both decision rules and predictions.
• Developed two visualization techniques: Feature Plot (visualizing feature topology) and Rule Plot (visualizing decision rules).
• Evaluated the approach using the "Glass" dataset and a user study.

Main Results:

• The proposed clustering and visualization methods effectively represent the collective behavior of random forests.
• Users could gain insights into model performance through the visualized clusters and individual tree structures.
• The new distance metric meaningfully groups similar decision trees.

Conclusions:

• The developed visualization system significantly enhances the interpretability of random forests.
• Clustering similar decision trees offers a scalable approach to understanding complex machine learning models.
• The Feature Plot and Rule Plot provide valuable tools for analyzing decision tree structures and rules.