Reproduce “Geometry- and Accuracy-Preserving Random Forest Proximities” on OpenResearch

About this paper

Jake S. Rhodes, Adele Cutler, Kevin R. Moon

Researchers at Utah State University introduce Random Forest-Geometry- and Accuracy-Preserving (RF-GAP) proximities, a novel measure that provably and empirically matches a random forest's out-of-bag predictions for both regression and classification. This new proximity measure leads to more accurate data visualizations that reflect learned geometry, improved imputation accuracy, and clearer outlier detection.

Key insights

• Random Forest predictions can be directly and provably reconstructed using a specific form of proximity-weighted nearest neighbors, revealing a deeper understanding of the forest's learned data geometry.
• Existing random forest proximity definitions, including the original, create a distorted view of the data geometry learned by the random forest, often over-emphasizing class separation or under-representing internal model structure.
• Aligning proximity measures with the random forest's predictive mechanism (OOB predictions) is crucial for improving the reliability and interpretability of all downstream applications that leverage these proximities.

Summary by alphaXiv.