Turbo-SMT: Parallel Coupled Sparse Matrix-Tensor Factorizations and Applications.

Area of Science:

• Computational neuroscience
• Machine learning
• Data science

Background:

• Correlating human brain neural activity with word properties presents a significant challenge.
• Existing Coupled Matrix-Tensor Factorization (CMTF) methods struggle with large datasets that exceed main memory capacity.

Purpose of the Study:

• To develop a scalable meta-method for enhancing any CMTF solver to handle large datasets.
• To improve the efficiency, interpretability, and parallelization of CMTF algorithms.
• To apply the enhanced method to neuroscience and social network datasets.

Main Methods:

• Introduction of Turbo-SMT, a meta-method designed to boost performance and scalability of existing CMTF algorithms.
• Improvements to the Alternating Least Squares (ALS) algorithm for enhanced efficiency and robustness.
• Application of Turbo-SMT to a (nouns, brain voxels, human subjects) tensor and a (nouns, properties) matrix dataset (BrainQ).

Main Results:

• Turbo-SMT significantly enhances CMTF algorithm performance, achieving up to 65-fold speedup through parallelization.
• The method produces sparse and interpretable latent variables, successfully predicting brain activity.
• Application to a Facebook dataset demonstrated Turbo-SMT's ability to identify anomalies like spammer activities.

Conclusions:

• Turbo-SMT offers a general and efficient solution for analyzing large-scale, multi-modal datasets using CMTF.
• The meta-method effectively extracts meaningful latent variables from complex data, with applications in neuroscience and anomaly detection.