PDPilot: Exploring Partial Dependence Plots Through Ranking, Filtering, and Clustering

Area of Science:

• Machine Learning
• Data Visualization
• Scientific Computing

Background:

• Partial dependence plots (PDPs) and individual conditional expectation (ICE) plots are crucial for interpreting machine learning (ML) models on tabular data.
• Analyzing these plots becomes challenging with a high number of features, hindering efficient model exploration.

Purpose of the Study:

• To develop and evaluate new techniques for ranking and filtering PDP and ICE plots.
• To enhance the efficiency of ML practitioners in exploring model behavior and identifying significant feature impacts.
• To integrate these novel techniques into a user-friendly visual analytics tool.

Main Methods:

• Development of novel algorithms for ranking and filtering PDP and ICE plots.
• Adaptation and integration of existing line clustering strategies for ICE plots.
• Implementation of these techniques within PDPilot, a visual analytics tool for Jupyter notebooks.
• Empirical study involving 7 ML practitioners to assess the usability of the developed techniques.

Main Results:

• The study presents new, effective techniques for prioritizing and selecting relevant PDP and ICE plots.
• The integration into PDPilot facilitates efficient exploration and analysis of ML model behavior.
• User study demonstrates the practical utility of the developed ranking, filtering, and clustering methods.

Conclusions:

• The developed techniques significantly improve the efficiency of analyzing ML models using PDP and ICE plots.
• PDPilot, with its integrated features, offers a valuable tool for ML practitioners.
• These advancements contribute to more interpretable and understandable machine learning models.