[ICML 2021] Part 3: Spatio-temporal Time-series Forecasting

Time series data is used in various fields such as climate change, economics, and healthcare, and research on Time series forecasting, which predicts future values or data distributions based on past data, is being actively conducted.

For future prediction, it is important to identify the Trend and Seasonality of the data. In particular, in a Multi-variate time-series environment where multiple types of time-series data are given, it is essential to identify Spatial correlation, the interrelationship between data.

Recently, Graph-based network structures have been widely used to model complex relationships between time series. This approach represents each Node as time-series data and Edges as relationships between the data.

This post covers two major papers introduced at ICML 2021:

• PAL (RNN with Particle Flow for Probabilistic Spatio-temporal Forecasting): Combines RNN with GCN (Graph Convolutional Network), and performs both point forecasting and probabilistic distribution of future data through Particle flow.
• Z-GCNETs (Time Zigzags at Graph Convolutional Networks for Time Series Forecasting): Proposes a time-series forecasting technique utilizing graph convolutional networks.