MOUNTAINEER: Topology-Driven Visual Analytics for Comparing Local Explanations.

Area of Science:

• Machine Learning
• Data Visualization
• Topological Data Analysis

Background:

• Increasing use of black-box Machine Learning (ML) models in critical applications necessitates transparency and accountability.
• Existing local explainability methods for ML models are difficult to evaluate and compare due to data complexity and method variability.
• Topological Data Analysis (TDA) offers a potential solution by transforming explanations into uniform graph representations for comparison.

Purpose of the Study:

• To present Mountaineer, a novel topology-driven visual analytics tool for analyzing and comparing ML explanations.
• To enable interactive exploration of ML explanations by linking topological graphs to data distributions, model predictions, and feature attributions.
• To facilitate deeper insights into explanation techniques and underlying data distributions for informed conclusions on model behavior.

Main Methods:

• Utilized Topological Data Analysis (TDA) to convert ML attributions into comparable graph representations.
• Developed Mountaineer, an interactive visual analytics tool for exploring these topological graphs.
• Linked topological graphs back to original data, model predictions, and feature attributions for comprehensive analysis.
• Conducted case studies with real-world data and expert interviews for evaluation.

Main Results:

• Demonstrated Mountaineer's capability to compare black-box ML explanations and identify causes of disagreement between methods.
• Showcased the tool's utility in comparing and understanding different ML models.
• Expert interviews confirmed the tool's effectiveness in gaining insights into ML explanations and model behavior.

Conclusions:

• Mountaineer provides a robust framework for evaluating and comparing ML explanations using TDA.
• The tool enhances transparency and accountability in ML by facilitating deeper understanding of model predictions.
• Mountaineer aids ML practitioners in making well-founded conclusions about model behavior and explanation techniques.