# Spatio-Temporal Graph Neural Point Process for Traffic Congestion Event Prediction

**Authors:** Guangyin Jin, Lingbo Liu, Fuxian Li, Jincai Huang  
**arXiv:** [2311.08635](https://arxiv.org/abs/2311.08635)  
**Date:** 2023-11-15  
**Summary page:** https://www.summarizepaper.com/en/paper/2311.08635v1/

## Key points

- Predicting traffic congestion events in intelligent transportation systems is crucial but complex
- Existing methods struggle with sparse congestion events over time
- A novel approach called Spatio-Temporal Graph Neural Point Process (STGNPP) is proposed
- STGNPP integrates a spatio-temporal graph learning module to capture long-range dependencies and extract hidden representations and event details
- Continuous gated recurrent unit models congestion evolution patterns, while periodic gated mechanism enhances intensity function calculation for improved accuracy
- Experimental results show that STGNPP outperforms existing approaches in predicting occurrence time and duration of future congestions

## Summary

, , , , 

In the realm of intelligent transportation systems, predicting traffic congestion events is a crucial yet complex task. Existing methods focus on forecasting dense variables using spatio-temporal graph-based neural networks, but struggle with sparse congestion events over time. To address this, a novel approach called Spatio-Temporal Graph Neural Point Process (STGNPP) is proposed. STGNPP integrates a spatio-temporal graph learning module to capture long-range dependencies and extract hidden representations and event details. A continuous gated recurrent unit models congestion evolution patterns, while a periodic gated mechanism enhances intensity function calculation for improved accuracy. The overall architecture includes road network data, historical traffic states, spatio-temporal indexes, and the Spatio-Temporal Graph Learning Module with link-wise Transformer and graph convolution layers. Experimental results show that STGNPP outperforms existing approaches in predicting both occurrence time and duration of future congestions by combining spatio-temporal graph learning with neural point process modeling for enhanced performance compared to traditional methods.