Probabilistic non-negative matrix factorization: theory and application to microarray data analysis.

Area of Science:

• Multivariate data analysis
• Machine learning
• Bioinformatics

Background:

• Non-negative matrix factorization (NMF) is vital for interpreting multivariate data, but its deterministic nature is challenged by real-world stochastic data and noise.
• The impact of noise on NMF's stability and convergence remains unclear, limiting its application in noisy environments.

Purpose of the Study:

• To generalize NMF for optimal non-negative factorization.
• To extend NMF to a probabilistic framework (PNMF) for handling data noise.
• To evaluate PNMF's performance in clustering and classifying DNA microarrays.

Main Methods:

• Generalized deterministic NMF with a broad class of update rules.
• Developed probabilistic NMF (PNMF) where the Maximum a posteriori (MAP) estimate solves a weighted regularized NMF problem.
• Derived update rules for PNMF convergence and applied it to DNA microarrays data.

Main Results:

• The generalized NMF ensures convergence towards optimal factorizations.
• PNMF effectively models data noise, with MAP estimates corresponding to a regularized NMF problem.
• PNMF demonstrated superior clustering stability and classification accuracy over deterministic and sparse NMF on DNA microarrays.

Conclusions:

• PNMF offers a robust probabilistic extension to NMF, adept at handling noisy, stochastic data.
• The developed PNMF framework provides improved performance in data clustering and classification tasks, particularly for biological data.
• This work advances NMF applications by incorporating probabilistic modeling for enhanced stability and accuracy.