Scalable non-negative matrix tri-factorization.

Area of Science:

• Computational Biology
• Data Science
• Machine Learning

Background:

• Matrix factorization is a key tool for pattern discovery in biomedical data analytics.
• Current algorithms face scalability challenges with large datasets due to computational intensity.
• Matrix tri-factorization offers a more flexible approach by inferring multiple latent spaces.

Purpose of the Study:

• To develop a scalable approach for matrix tri-factorization.
• To address the computational challenges of analyzing large biomedical datasets.
• To enhance feature discovery and data integration capabilities.

Main Methods:

• A novel block-wise strategy for latent factor learning in matrix tri-factorization was developed.
• The data matrix is partitioned into submatrices for parallel processing.
• The approach was implemented and tested on multi-processor and multi-GPU architectures.

Main Results:

• The block-wise approach demonstrates mathematical equivalence to serial matrix tri-factorization.
• Significant speedups were observed on large biomedical datasets using multi-GPU systems.
• Performance improvements exceeded 100-times faster compared to single-processor methods on a four-GPU system.

Conclusions:

• A general method for scaling non-negative matrix tri-factorization has been successfully proposed.
• The approach is particularly effective for parallel matrix factorization in multi-GPU environments.
• This method is expected to advance latent factor analysis, especially for large-scale data integration tasks.