Reproduce “Density-Aware Farthest Point Sampling” on OpenResearch

About this paper

Paolo Climaco, Jochen Garcke

A new sampling algorithm, Density-Aware Farthest Point Sampling (DA-FPS), enhances the average prediction performance of Lipschitz continuous regression models by employing a density-aware weighted fill distance for passive and model-agnostic data selection. This strategy, developed by researchers at the University of Bonn, UCLA, and Fraunhofer SCAI, consistently reduced the mean absolute error (MAE) compared to baselines on molecular property prediction and other datasets, especially with larger training sets.

Key insights

• The expected prediction error of Lipschitz continuous regression models is linearly bounded by a weighted fill distance, indicating that both geometric coverage and data distribution alignment are critical for accuracy.
• A simplified weight function derived from kNN density estimates can effectively approximate the ideal density ratio, circumventing issues with direct ratio calculations.
• An initial phase of uniform point selection, similar to standard FPS, enhances the reliability of density estimates and ensures broad coverage, which is beneficial before applying density-aware weighting.

Summary by alphaXiv.