Mixed Membership sub-Gaussian Models
New spectral algorithm beats Gaussian mixture limits for overlapping data.
Huan Qing introduces the mixed membership sub-Gaussian model, a novel extension of the Gaussian mixture model that addresses a fundamental limitation: classical models force each observation into exactly one component. This new approach allows observations to have partial membership in multiple latent components, making it suitable for real-world data with overlapping structures, such as genetic profiles, social network communities, or mixed-topic documents.
The paper presents an efficient spectral algorithm for estimating mixed membership vectors, proving that under mild separation conditions on component centers, the estimation error can be made arbitrarily small with high probability. This is the first computationally efficient estimator with such a vanishing-error guarantee for a mixed-membership Gaussian mixture model. Extensive experiments demonstrate superior performance over traditional methods that ignore mixed memberships, with 30 pages, 6 figures, and 2 tables supporting the findings.
- The model extends Gaussian mixtures to allow each observation to belong to multiple components, addressing overlapping structures in genetics, social networks, and text mining.
- A spectral algorithm estimates per-individual membership vectors with arbitrarily small error under mild separation conditions, a first for this model type.
- Extensive experiments show the method outperforms existing approaches that ignore mixed memberships, with 30 pages of analysis.
Why It Matters
Enables accurate clustering of overlapping data, improving analysis in genetics, social networks, and text mining.