Bayesian Matrix Factorization for Semibounded Data.

Area of Science:

• Machine Learning
• Statistical Modeling
• Data Analysis

Background:

• Bayesian non-negative matrix factorization (BNMF) is a common technique.
• Existing methods may not be optimal for semibounded data.
• There is a need for scalable and adaptable factorization methods.

Purpose of the Study:

• Propose a novel BNMF model for semibounded data.
• Develop a variational Bayesian inference approach for parameter estimation.
• Introduce an online extension for enhanced scalability and streaming data adaptation.

Main Methods:

• Developed a BNMF model assuming Inverted Beta distribution for semibounded data.
• Utilized variational Bayesian inference for parameter estimation.
• Derived an analytically tractable solution using a lower bound approximation of the objective function.
• Proposed an online extension for scalability and streaming data.

Main Results:

• The proposed BNMF model effectively handles semibounded data.
• Variational Bayesian inference provides an efficient estimation method.
• The online extension demonstrates improved scalability for large datasets.
• The model shows strong performance across diverse applications.

Conclusions:

• The novel BNMF technique offers a robust solution for semibounded data factorization.
• The variational Bayesian inference and online extension enhance practical applicability.
• The model's versatility is confirmed through evaluations on multiple real-world applications.