Probabilistic Matrix Factorization for Data With Attributes Based on Finite Mixture Modeling.

Area of Science:

• Machine Learning
• Statistical Modeling
• Data Mining

Background:

• Matrix factorization (MF) is a common technique for dimensionality reduction.
• Existing probabilistic MF models can be computationally intensive.
• Joint modeling of data and attribute matrices is an active research area.

Purpose of the Study:

• To propose a novel generative latent variable model for matrix factorization.
• To jointly model data and attribute matrices for improved performance.
• To develop a computationally efficient inference method.

Main Methods:

• A generative latent variable model is proposed, assuming Gaussian entries with means derived from inner products of factor matrices.
• Finite mixture of Gaussians is used as the prior for factor matrix columns.
• Variational Bayes inference is employed for efficient computation of posteriors and model parameters.

Main Results:

• The proposed model demonstrates computational efficiency compared to sampling-based probabilistic MF models.
• Experimental results show state-of-the-art performance in collaborative filtering tasks.
• The model also achieves superior results in community detection tasks.

Conclusions:

• The proposed generative latent variable model offers an efficient and effective approach to matrix factorization.
• Joint modeling of data and attribute matrices enhances performance in various applications.
• The variational Bayes inference provides a computationally advantageous alternative to existing methods.