Researchers propose Wasserstein Filtering to clean AI training data

A new arXiv paper introduces Wasserstein Filtering (WF), a sample selection method for robust distribution learning. WF discards suspicious samples and estimates the clean population distribution from the remaining data, selecting a subset whose empirical distribution maximizes its Wasserstein distance to the fully contaminated empirical distribution to isolate geometrically influential outliers. The paper proposes three algorithms—SinkMarg, SinkWF, and SlicedWF—using entropic optimal transport and sliced Wasserstein approximations. It also introduces the Far Exclusion and Local Projection (FELP) contamination model and proves that the WF estimator achieves minimax optimality over distribution families with bounded covariance. Experiments on synthetic datasets, benchmark anomaly detection suites, and robust generative learning with diffusion models show that WF serves as a practical, model-agnostic preprocessing tool, delivering competitive outlier detection performance and substantial downstream benefits for generative modeling under heavy contamination.