Cutstats Method for Noisy Data Subset Selection Enhanced by Data Symmetries
A new study on arXiv (2605.01874) formally establishes that the performance of cutstats, a method for selecting low-noise subsets of training data, depends on k-nearest neighbors (k-NN) accuracy. The research shows that exploiting data invariance and underlying symmetries can significantly improve k-NN performance in high-dimensional noisy environments, bringing it closer to the Bayes optimal classifier. The work addresses the challenge of label noise in large datasets collected from diverse sources, where optimal subsets can yield performance comparable to noise-free training.
Key facts
- arXiv paper 2605.01874 analyzes cutstats for noisy data subset selection.
- Cutstats uses k-nearest neighbors (k-NN) to detect low-noise samples.
- Performance of cutstats depends on k-NN accuracy.
- Data invariance and symmetries can enhance k-NN in high dimensions.
- Improved k-NN approaches Bayes optimal classifier under label noise.
- Label noise arises from diverse data sources.
- Optimal subsets can match noise-free training performance.
- Study focuses on high-dimensional data performance.
Entities
Institutions
- arXiv