Embedding spatial context in urban traffic forecasting with contrastive pre-training

TL;DR

The paper addresses improving urban traffic forecasting by incorporating spatial context derived from road topology and OpenStreetMap features through a traffic quotient graph. It introduces a geometric encoder trained with contrastive learning on augmented quotient graphs and integrates it with existing spatio-temporal forecasting models via spatially gated addition, enabling improvements without extra traffic data. Key contributions include the traffic quotient graph construction, OSM-based feature extraction, a contrastive pre-training pipeline for the geometric encoder, and empirical demonstrations of enhanced generalization and accuracy across multiple datasets. This approach broadens the utility of non-traffic contextual data for cyber-physical transport systems and offers practical avenues for richer, context-aware forecasting.

Abstract

Urban traffic forecasting is a commonly encountered problem, with wide-ranging applications in fields such as urban planning, civil engineering and transport. In this paper, we study the enhancement of traffic forecasting with pre-training, focusing on spatio-temporal graph methods. While various machine learning methods to solve traffic forecasting problems have been explored and extensively studied, there is a gap of a more contextual approach: studying how relevant non-traffic data can improve prediction performance on traffic forecasting problems. We call this data spatial context. We introduce a novel method of combining road and traffic information through the notion of a traffic quotient graph, a quotient graph formed from road geometry and traffic sensors. We also define a way to encode this relationship in the form of a geometric encoder, pre-trained using contrastive learning methods and enhanced with OpenStreetMap data. We introduce and discuss ways to integrate this geometric encoder with existing graph neural network (GNN)-based traffic forecasting models, using a contrastive pre-training paradigm. We demonstrate the potential for this hybrid model to improve generalisation and performance with zero additional traffic data. Code for this paper is available at https://github.com/mattchrlw/forecasting-on-new-roads.

Figures (10)

• Figure 1: Our geometric encoder pre-training method constructs traffic quotient graphs, generates stochastic pairs of graphs augmented with real OpenStreetMap contextual data, and feeds them into a geometric encoder. This figure illustrates this process applied to the METR-LA dataset.
• Figure 2: A diagram of the various components of the paper, with the contribution areas highlighted in yellow and the application area in green.
• Figure 3: The METR-LA and PEMS-BAY datasets 10.1007/s10489-021-02648-0.
• Figure 4: Left: the traffic data $T = (T_V, T_E)$. Right: the road graph $R = (R_V, R_E)$.
• Figure 5: The constructed traffic quotient graphs for two different datasets. Left: METR-LA. Right: PEMS-BAY. Note that for METR-LA the resultant traffic quotient graph has an isolated node, but when using Graph WaveNet connections can still be learned using the self-adaptive adjacency matrix.