Feature interaction graphs for exact interpretable learning solver selection: an empirical diagnostic study.

Area of Science:

• Machine Learning
• Explainable AI (XAI)

Background:

• Exact interpretable learning is crucial for regulated decision-making.
• Solver runtime variability across datasets and families poses a challenge.

Purpose of the Study:

• Introduce structural meta-features from Feature Interaction Graphs (FIGs) for interpretable solver selection.
• Develop a transparent decision-tree selector using these FIG features.

Main Methods:

• Construct FIGs from binarized tabular data using pairwise mutual information.
• Extract topology-aware signatures like density and estimated treewidth.
• Utilize a shallow decision-tree for meta-feature-based solver selection.

Main Results:

• FIG features provide an interpretable structural view of solver behavior.
• Topology-aware profiling highlights meaningful dataset variations.
• While benchmark saturation limits routing gains, FIGs prove effective on harder instances.

Conclusions:

• FIGs serve as a principled, interpretable diagnostic tool for algorithm selection in exact learning.
• The diagnostic relevance of FIGs is most apparent on instances with substantial solver runtime differences.