Reproduce “Woodelf++: A Fast and Unified Partial Dependence Plot Algorithm for Decision Tree Ensembles” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Ron Wettenstein, Alexander Nadel, Udi Boker
WOODELF++ is a unified algorithm developed by researchers at Reichman University and Technion that significantly enhances the computational efficiency of Partial Dependence Plots (PDPs), Joint-PDPs, and Any-Order Partial Dependence Interaction Values (PDIVs) for decision tree ensembles. The algorithm achieves substantial speedups, converting previously intractable tasks (e.g., millions of years for Any-Order-PDIVs) into computations completed in seconds or minutes, and introduces the concept of exact "Full PDPs."
Key insights
- The definition of Partial Dependence Values can be formally mapped to the average model prediction within a cooperative game where specific features are fixed and others are marginalized, enabling a unified algorithmic approach.
- Leveraging the linearity of Pseudo Boolean functions represented in Weighted Disjunctive Normal Form allows for the decomposition of complex global interpretability metrics into efficiently computable contributions from individual logical cubes.
- By moving from sampling to an exact enumeration of relevant split points, "Full PDPs" provide a precise and complete visualization of feature dependence, overcoming limitations of traditional sampled plots.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.