Tensor Completion via Completable Substructure Sampling for Online Spatiotemporal Data Analysis
Abstract
Tensor representation of data captures inter-dependencies along multiple directions, which provides a natural framework for noncommutative analysis of interdependent multimodal data. However, the specific tensor theory for noncommutative analysis is yet to be established. The objective of this project is to create a new online tensor-learning approach for large-scale spatial-temporal data stream analysis. Focus will be placed on developing effective online tensor completion techniques by leveraging recent successes in machine learning, numerical analysis and statistics. This approach relaxes the incoherent assumptions common in existing approaches and explores the completable substructures of tensors. It examines the local structures of a tensor (i.e., sub-tensors) via random matrix theories and enables localized update via distributed computing to achieve both prediction improvement and computational efficiency. The proposed solution is nonparametric so that it can relax the strong model assumptions in previous work and fully utilize the large amount of data available.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Jan 12, 2017
- Source ID
- W911NF1510491
Entities
People
- Yan Y Liu
Organizations
- Army Contracting Command
- United States Army
- University of Southern California