A Spectral Method for Identifiable Grade of Membership Analysis with Binary Responses.

Area of Science:

• Statistics
• Data Analysis
• Machine Learning

Background:

• Grade of Membership (GoM) models are advanced individual-level mixture models for multivariate categorical data.
• GoM allows subjects to have mixed memberships in latent profiles, offering richer modeling than latent class models.
• However, GoM models present significant identifiability and estimation challenges.

Purpose of the Study:

• To propose a singular value decomposition (SVD)-based spectral approach for Grade of Membership (GoM) analysis.
• To address the identifiability and estimation challenges inherent in GoM models for multivariate binary responses.
• To develop a computationally efficient and scalable method for GoM analysis.

Main Methods:

• Utilized singular value decomposition (SVD) by leveraging the low-rank decomposition of the data matrix expectation under a GoM model.
• Developed conditions for expectation identifiability and extracted leading singular vectors for parameter estimation.
• Established estimator consistency in a double-asymptotic regime (increasing subjects and items).

Main Results:

• The proposed spectral method demonstrates superior efficiency and accuracy compared to traditional Bayesian or likelihood-based methods.
• The method is computationally advantageous and scalable for large-scale, high-dimensional data.
• Successful application of the method to a personality test dataset validated its practical utility.

Conclusions:

• The SVD-based spectral approach provides a powerful and efficient new tool for Grade of Membership (GoM) model analysis.
• This method overcomes computational limitations and scalability issues of existing techniques.
• The approach is effective for analyzing mixed memberships in multivariate categorical data, with broad applicability.