TreePOD: Sensitivity-Aware Selection of Pareto-Optimal Decision Trees

Area of Science:

• Computer Science
• Machine Learning
• Data Mining

Background:

• Automating decision tree selection is challenging due to the need for domain expertise and model purpose understanding.
• Balancing accuracy with interpretability in decision trees is a critical, yet difficult, objective.

Purpose of the Study:

• To introduce TreePOD, a novel approach for sensitivity-aware model selection in decision trees.
• To facilitate automated identification of decision trees that balance multiple objectives, such as accuracy and interpretability.

Main Methods:

• TreePOD explores a large set of candidate decision trees generated by sampling tree construction algorithm parameters.
• Utilizes visualizations of quantitative and qualitative tree aspects to provide an overview of characteristics.
• Employs Pareto-optimal candidates for efficient selection guidance along trade-offs and full-factorial sampling for sensitivity analysis.

Main Results:

• Demonstrates TreePOD's support for various decision tree selection tasks and its integration into a holistic workflow.
• Presents a case study on predicting critical power grid states, validating the approach.
• Reports positive qualitative feedback from domain experts, indicating enhanced user confidence and efficiency.

Conclusions:

• TreePOD enables users, regardless of statistical background, to confidently and efficiently identify suitable decision trees.
• The approach effectively addresses the challenge of balancing multiple objectives in decision tree construction and selection.
• TreePOD facilitates a more automated and guided process for decision tree model selection.