Text / embedding outlier screening

Embedding spaces often contain topical clusters and occasional off-topic documents. Robust scatter estimators provide a simple way to rank unusual embeddings without training a supervised classifier.

Result at a glance

The example selects StudentTScatter and recovers all injected outlier embeddings in the lightweight simulation. This makes a good entry point for document review, moderation, and search-quality diagnostics.

What the data represent

The data are synthetic embedding-like vectors: a central topic cloud plus a small group of off-topic points. The goal is to mimic the geometry of sentence or document embeddings without requiring external models.

Why this estimator

AutoRobustScatter chooses among robust scatter candidates. Student-t scatter is often a good compromise for diffuse, heavy-tailed embedding clouds.

Reproduce the result

python examples/use_case_text_embedding_outliers.py

Output from the run

embedding outlier screening
selected=StudentTScatter
precision=1.000, recall=1.000, detected=55
saved diagnostics to results/use_cases/embedding_outliers

Figures and diagnostics

How to read the result

Use the distance panel as a ranked review queue. The top-scoring embeddings are candidates for off-topic or low-quality items; the threshold should usually be calibrated by review capacity.

What this does not prove

Real embeddings can be strongly multimodal. For multiple legitimate topics, prefer cluster-aware robust distances or segment the corpus before fitting.