Forecasting Multivariate Time Series with Nonstationary Temporal Matrix Factorization
High-dimensional and sparse time series forecasting on Uber movement speed data
Uber movement project (http://movement.uber.com/) provides data and tools for cities to more deeply understand and address urban transportation challenges. Uber movement speed data measure hourly street speeds across a city (e.g., New York City, Seattle, London) to enable data-driven city planning and decision making. These data are indeed multivariate time series with N road segments and T time steps (hours), and are featured as high-dimensional, sparse, and nonstationary. To overcome the challenge created by these complicated data behaviors, we propose a nonstationary temporal matrix factorization (NoTMF) framework for multivariate time series forecasting on high-dimensional and sparse Uber movement speed data.
More about this story:
- The adapted data sets and Python codes: https://github.com/xinychen/tracebase
- Another relevant blog post (Temporal matrix factorization for multivariate time series forecasting): https://medium.com/p/b1c59faf05ea
- Slides: https://xinychen.github.io/slides/notmf.pdf
Motivation
Real-world traffic speed data on urban road networks are high-dimensional, sparse, and nonstationary. But it is not hard to deal with these data behaviors. We consider to…