Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Heterogeneity-Informed Meta-Parameter Learning for Spatiotemporal Time Series Forecasting

Zheng Dong, Renhe Jiang, Haotian Gao, Hangchen Liu, Jinliang Deng, Qingsong Wen, Xuan Song

TL;DR

A novel Heterogeneity-Informed Spatiotemporal Meta-Network (HimNet) is developed for spatiotemporal time series forecasting that achieves state-of-the-art performance while exhibiting superior interpretability.

Abstract

Spatiotemporal time series forecasting plays a key role in a wide range of real-world applications. While significant progress has been made in this area, fully capturing and leveraging spatiotemporal heterogeneity remains a fundamental challenge. Therefore, we propose a novel Heterogeneity-Informed Meta-Parameter Learning scheme. Specifically, our approach implicitly captures spatiotemporal heterogeneity through learning spatial and temporal embeddings, which can be viewed as a clustering process. Then, a novel spatiotemporal meta-parameter learning paradigm is proposed to learn spatiotemporal-specific parameters from meta-parameter pools, which is informed by the captured heterogeneity. Based on these ideas, we develop a Heterogeneity-Informed Spatiotemporal Meta-Network (HimNet) for spatiotemporal time series forecasting. Extensive experiments on five widely-used benchmarks demonstrate our method achieves state-of-the-art performance while exhibiting superior interpretability. Our code is available at https://github.com/XDZhelheim/HimNet.

Heterogeneity-Informed Meta-Parameter Learning for Spatiotemporal Time Series Forecasting

TL;DR

A novel Heterogeneity-Informed Spatiotemporal Meta-Network (HimNet) is developed for spatiotemporal time series forecasting that achieves state-of-the-art performance while exhibiting superior interpretability.

Abstract

Spatiotemporal time series forecasting plays a key role in a wide range of real-world applications. While significant progress has been made in this area, fully capturing and leveraging spatiotemporal heterogeneity remains a fundamental challenge. Therefore, we propose a novel Heterogeneity-Informed Meta-Parameter Learning scheme. Specifically, our approach implicitly captures spatiotemporal heterogeneity through learning spatial and temporal embeddings, which can be viewed as a clustering process. Then, a novel spatiotemporal meta-parameter learning paradigm is proposed to learn spatiotemporal-specific parameters from meta-parameter pools, which is informed by the captured heterogeneity. Based on these ideas, we develop a Heterogeneity-Informed Spatiotemporal Meta-Network (HimNet) for spatiotemporal time series forecasting. Extensive experiments on five widely-used benchmarks demonstrate our method achieves state-of-the-art performance while exhibiting superior interpretability. Our code is available at https://github.com/XDZhelheim/HimNet.
Paper Structure (21 sections, 12 equations, 7 figures, 7 tables)

This paper contains 21 sections, 12 equations, 7 figures, 7 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Spatiotemporal Forecasting
  4. Meta-Parameter Learning
  5. Problem Definition
  6. Methodology
  7. Spatiotemporal Heterogeneity Modeling
  8. Spatiotemporal Meta-Parameter Learning
  9. Heterogeneity-Informed Spatiotemporal Meta-Network
  10. Experiment
  11. Experimental Setup
  12. Performance Evaluation
  13. Ablation Study
  14. Hyper-Parameter Study
  15. Efficiency Study
  16. ...and 6 more sections

Figures (7)

  • Figure 1: Spatiotemporal heterogeneity. (a): Locations of two selected areas. $A_1$ is in the downtown region, while $A_2$ is a residential area. (b): Illustration of spatial heterogeneity, with differing vehicle speed distributions between $A_1$ and $A_2$. (c): Demonstration of temporal heterogeneity, showing distinct patterns between different time periods.
  • Figure 2: The overall illustration of the proposed Heterogeneity-Informed Meta-Parameter Learning and HimNet model.
  • Figure 3: RMSE w.r.t. embedding dimensions on METRLA.
  • Figure 4: t-SNE visualization of the temporal and spatial meta-parameters. (a): Visualization for each day in a week where every cluster contains 24 points representing each hour's temporal meta-parameter. (b): The cosine similarity matrix of the meta-parameters across 24 hours in a day. (c): Visualization for each sensor's spatial meta-parameter in METRLA dataset. (d): Sensor points in (c) plotted on map, with only one lane of points shown for dual carriageways. (e): Clusters 3 and 13 are located on opposite lanes of the same road. (e): The one-week time series of clusters 3, 10, and 13.
  • Figure 5: The evolving ST-mixed meta-parameters. (a): Distribution of clusters 3 and 6 around a crossroad area. (b): The traffic speed time series of the two clusters at peak hour and off-peak hour. (c): A closer view of 10 sensors at the crossroad. (d): The cosine similarity of their ST-mixed meta-parameters across three time periods.
  • ...and 2 more figures