Bayesian Temporal Factorization for Multidimensional Time Series Prediction.

Area of Science:

• Data Science
• Machine Learning
• Time Series Analysis

Background:

• Spatiotemporal data is increasingly common in applications like urban monitoring.
• Predicting from large, high-dimensional time series with missing data is challenging.

Purpose of the Study:

• To propose a novel Bayesian temporal factorization (BTF) framework.
• To model multidimensional time series, especially spatiotemporal data, with missing values.
• To enable accurate predictions and uncertainty quantification.

Main Methods:

• Integrated low-rank matrix/tensor factorization with vector autoregressive (VAR) processes.
• Developed a probabilistic graphical model for global and local consistency.
• Employed efficient Gibbs sampling for inference and real-time updating.

Main Results:

• The BTF framework effectively models spatiotemporal data with missing values.
• Probabilistic predictions and uncertainty estimates were generated without imputation.
• Demonstrated superior performance over existing methods in imputation and prediction tasks.

Conclusions:

• The proposed BTF framework offers a robust solution for spatiotemporal data analysis.
• It successfully handles missing data and provides reliable predictions.
• BTF advances the state-of-the-art in large-scale time series modeling.