Soft clustering assigns every observation a score for each cluster rather than placing it in exactly one cluster. Gaussian mixture models use these scores to express how well a record matches different multivariate Gaussian processes. section 5.6
Overlapping clusters represent observations plausibly generated by multiple processes. For outlier detection, records far from all substantial Gaussian centers are unlikely, while cluster-specific structure can outperform a single covariance model on multi-cluster data. section 5.6
Gaussian mixtures for outlier detection
A scikit-learn example creates three dense synthetic clusters and fifteen dispersed points, fits Gaussian mixtures with candidate component counts, and identifies outliers from the resulting model.
