S3CMTF: Fast, accurate, and scalable method for incomplete coupled matrix-tensor factorization.

Area of Science:

• Data Mining
• Machine Learning
• Scientific Computing

Background:

• Coupled matrix-tensor factorization (CMTF) is vital for extracting relations from combined matrix and tensor data.
• Existing CMTF methods struggle with accuracy, speed, and scalability for large, real-world datasets.
• The increasing size and dimensionality of data necessitate more efficient CMTF solutions.

Purpose of the Study:

• To develop a fast, accurate, and scalable CMTF method capable of handling large sparse tensors.
• To enable parallel processing for CMTF, overcoming limitations of existing non-parallelizable methods.
• To improve the efficiency and performance of relation extraction from coupled matrix and tensor data.

Main Methods:

• Proposed S3CMTF, a parallel sparse CMTF method utilizing derived gradient update rules.
• Implemented asynchronous partial gradient updates to avoid locking and enhance speed.
• Incorporated optimized storage and reuse of intermediate computations for performance boosts.

Main Results:

• S3CMTF demonstrates theoretical and empirical convergence to quality solutions.
• Achieved significant speedups, up to 930x faster than existing methods, with superior accuracy.
• Exhibited linear scalability with respect to data entries and processing cores.
• Successfully applied to Yelp rating data for pattern discovery.

Conclusions:

• S3CMTF represents a breakthrough in efficient and scalable CMTF.
• The method significantly advances the state-of-the-art in extracting hidden relations from complex datasets.
• S3CMTF offers a powerful tool for analyzing large-scale coupled matrix and tensor data in various domains.