Extracting trends from spatio-temporal data, such as the Google COVID-19 Search Trends Symptoms Dataset or Chicago Crime Dataset, can be used to investigate changes related to health or the environment, respectively. Tensor factorization can naturally capture space and time dependence to identify meaningful patterns.
Recent advances in federated tensor learning have further enabled joint learning across multiple sources in a privacy-preserving manner. Yet, measurements can be erroneous and missing in spatio-temporal data and can negatively impact the cur- rent federated tensor factorization approaches. In this paper, we develop a robust federated tensor factorization framework, FedTefid, that is not only efficient from a communication and computational perspective but also able to extract temporal patterns from incomplete data via a temporal smoothness constraint. Experiment re- sults show that our proposed method can recover the spatio-temporal patterns even with 90% of the measurements missing.
Table of Contents
Introduction Background Tensor Rank-one tensors Tensor matricization CP decomposition Generalized CP CP Decomposition with missing data Federated learning Federated tensor factorization Proposed method Federated GTF with missing data Gradient updates with observed entries Temporal smoothing L2 norm L1 norm FedTefid Experiments Datasets Baselines Hyper-parameter tuning Rank Smoothness penalty Uniformly missing data Skewed missing data Case study on COVID-19 Open Data Conclusion Bibliography
About this Honors Thesis
|Committee Chair / Thesis Advisor|
|Understanding spatial-temporal trends using communication-efficient federated tensor factorization for incomplete data ()||2022-05-16 17:43:46 -0400||