A data-diversity score flags odd samples without heavy tuning
A simple method spots unusual samples by how they shift data relationships.
- Publication
- arXiv
- Stage
- Preprint
- What we read
- Summary of the paper
- Authors
- Amey Pasarkar, Adji Bousso Dieng
- Universities and research institutions
- Lewis-Sigler Institute for Integrative Genomics, Princeton University, Department of Computer Science, Princeton University
Institution metadata: Original paper ↗
What they did and found
A new approach checks how removing one data point changes how the data relate to each other. It uses vector features that summarize data and avoids complex parameter tweaks. The method says it works fast across many kinds of data, including text and pictures.
Why it matters
For teams, this may help clean big data and find rare groups that could distort results, without fixed zones or tuned knobs.
The idea checks how much a point changes overall diversity when it is left out. It ranks points by disruption instead of fitting a model to every dataset.
In practice, this can help catch mislabeled items before building or updating systems, across data like text or pictures. It may miss anomalies if the summary features hide the problem.
What remains uncertain
The approach relies on the quality of the feature space and the chosen similarity rule. It may miss anomalies that look normal in the summary features and has been tested mainly on early data.
Original sources · 1
- Vendi Anomaly Scores for Efficient and Accurate Anomaly Detection ↗arXiv · 2026-11-09
Check the original paper for its authors, methods, version and access terms.
What is your take?
Ask a question, add useful context or share a different perspective. Keep the conversation respectful and grounded.