Reproduce “Machine Unlearning for Random Forests” 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
Jonathan Brophy, Daniel Lowd
Jonathan Brophy and Daniel Lowd from the University of Oregon developed Data Removal-Enabled (DaRE) forests, a Random Forest variant for efficient and exact removal of training data instances. DaRE forests achieve geometric mean speedups of 257x to over 1,272x compared to naive retraining for random data deletions, with minimal impact on predictive performance.
Key insights
- Efficient exact unlearning for Random Forests is achievable by selectively retraining only the minimal affected subtrees, rather than rebuilding the entire model.
- Incorporating "random" split nodes in the upper tree layers can significantly boost deletion efficiency by reducing retraining frequency, with a controllable and graceful impact on predictive performance.
- Omitting bootstrapping during Random Forest training can simplify exact unlearning without significantly impacting predictive performance, as it complicates instance removal.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.