Sparse data embedding and prediction by tropical matrix factorization.

Area of Science:

• Computational Biology
• Data Science
• Machine Learning

Background:

• Traditional matrix factorization models are linear and struggle with complex data relationships.
• Sparse data presents challenges for existing matrix factorization techniques.
• Non-negative Matrix Factorization (NMF) has limitations in capturing intricate patterns and extreme values.

Purpose of the Study:

• To introduce non-linearity into matrix factorization models using tropical semirings.
• To propose and evaluate a novel method, Sparse Tropical Matrix Factorization (STMF), for estimating missing values in sparse datasets.
• To compare the performance of STMF against NMF on both synthetic and real-world biological data.

Main Methods:

• Development of the Sparse Tropical Matrix Factorization (STMF) method.
• Utilizing tropical semirings to incorporate non-linearity into matrix factorization.
• Application and evaluation on synthetic datasets and gene expression data from The Cancer Genome Atlas (TCGA).

Main Results:

• STMF demonstrated higher correlation compared to NMF on synthetic data.
• STMF outperformed NMF on six out of nine gene expression datasets.
• STMF showed a superior ability to fit extreme values and distributions compared to NMF, which tends towards the mean.

Conclusions:

• STMF is the pioneering application of tropical semirings to sparse data.
• The study highlights the utility of semirings in handling data structures distinct from standard linear algebra.
• STMF offers a more effective approach for analyzing sparse biological data, particularly gene expression patterns.