Local Feature Relevance in Full-Covariance Gaussian Mixture Models via Penalized Classification EM
Year: 2026
Type: preprint
Abstract: In clustering by mixture models, relevant features are often cluster-specific and correlated within each group. Existing methods force an uncomfortable compromise: they either assume conditional independence to select features locally, or retain full covariances at the expense of a single global feature subset. We resolve this dilemma with a Full-Covariance Gaussian Mixture Model with Local Feature Relevance (FC-GMM-LFR). Each observation is governed by its own binary activation profile: active features follow the cluster's full-covariance Gaussian distribution, inactive features follow a shared background distribution, and a null profile naturally isolates non-informative observations. Because feature relevance is not identifiable from the likelihood alone, we select cluster-specific supports through a penalized classification criterion. We maximize this criterion with a monotone block coordinate ascent algorithm that classifies individual profiles, updates full covariances via a missing-data EM step under spectral regularization, and alternates with background estimation. In simulations, our method reliably recovers local active structures and outperforms competitors when clusters differ by their correlations. On sorted single-cell blood data, it accurately separates cell populations (ARI = 0.96), and its estimated activation rates closely reproduce the detection rates of marker genes.
Source: Research Square
Authors: Wilson Toussile, Siméon Fotso
Institution: Université de Yaoundé I
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7234883028
