Time series recovery from partial observations via Nonnegative Matrix Factorization
Abstract: In the modern analysis of time series, one may want to forecast several hundred or thousands of times series with possible missing entries. We introduce a novel algorithm to address these issues, referred to as Sliding Mask Method (SMM). SMM is a method based on the framework of predicting a time window and using completion of nonnegative matrices. This new procedure combines nonnegative factorization and matrix completion with hidden values (i.e., a partially observed matrix). From a theoretical point of view, we prove the statistical guarantees on uniqueness and robustness of the solutions of the completion of partially observed nonnegative matrices. From a numerical point of view, we present experiments on real-world and synthetic data-set confirm forecasting accuracy for the novel methodology.
Submission Length: Regular submission (no more than 12 pages of main content)
Previous TMLR Submission Url: https://openreview.net/forum?id=hI8U2hgfff
Changes Since Last Submission: The submission was desk rejected for 11pt format. We apologize for this inconvenience. The format is now 10pt as required.
Sincerely,
Assigned Action Editor: ~Seungjin_Choi1
Submission Number: 2738
Loading