Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

MoGERNN: An Inductive Traffic Predictor for Unobserved Locations in Dynamic Sensing Networks

Qishen Zhou, Yifan Zhang, Michail A. Makridis, Anastasios Kouvelas, Yibing Wang, Simon Hu

TL;DR

MoGERNN tackles the Forecasting Unobserved Nodes Under Dynamic Sensing networks (FUNDS) problem by combining a Mixture of Graph Experts (MoGE) block for initializing unobserved-node embeddings with a GRU-based Graph Encoder-Decoder that captures spatio-temporal dynamics. The model is trained inductively via random masking to simulate sensor changes and uses a sparse gating mechanism to fuse diverse spatial aggregators, enabling accurate predictions at both observed and unobserved locations without retraining when sensor configurations change. Empirical results on METR-LA and PEMS-BAY show MoGERNN consistently outperforms baselines, exhibits strong zero-shot performance for newly added sensors, and maintains competitive accuracy under dynamic sensor networks, highlighting its practical utility for adaptive traffic management. The work advances inductive spatio-temporal modeling in dynamic sensing contexts and suggests future directions in uncertainty estimation, multi-source data fusion, and multivariate traffic data integration.

Abstract

Given a partially observed road network, how can we predict the traffic state of unobserved locations? While deep learning approaches show exceptional performance in traffic prediction, most assume sensors at all locations of interest, which is impractical due to financial constraints. Furthermore, these methods typically require costly retraining when sensor configurations change. We propose MoGERNN, an inductive spatio-temporal graph representation model, to address these challenges. Inspired by the Mixture of Experts approach in Large Language Models, we introduce a Mixture of Graph Expert (MoGE) block to model complex spatial dependencies through multiple graph message aggregators and a sparse gating network. This block estimates initial states for unobserved locations, which are then processed by a GRU-based Encoder-Decoder that integrates a graph message aggregator to capture spatio-temporal dependencies and predict future states. Experiments on two real-world datasets show MoGERNN consistently outperforms baseline methods for both observed and unobserved locations. MoGERNN can accurately predict congestion evolution even in areas without sensors, offering valuable information for traffic management. Moreover, MoGERNN is adaptable to dynamic sensing networks, maintaining competitive performance even compared to its retrained counterpart. Tests with different numbers of available sensors confirm its consistent superiority, and ablation studies validate the effectiveness of its key modules.

MoGERNN: An Inductive Traffic Predictor for Unobserved Locations in Dynamic Sensing Networks

TL;DR

MoGERNN tackles the Forecasting Unobserved Nodes Under Dynamic Sensing networks (FUNDS) problem by combining a Mixture of Graph Experts (MoGE) block for initializing unobserved-node embeddings with a GRU-based Graph Encoder-Decoder that captures spatio-temporal dynamics. The model is trained inductively via random masking to simulate sensor changes and uses a sparse gating mechanism to fuse diverse spatial aggregators, enabling accurate predictions at both observed and unobserved locations without retraining when sensor configurations change. Empirical results on METR-LA and PEMS-BAY show MoGERNN consistently outperforms baselines, exhibits strong zero-shot performance for newly added sensors, and maintains competitive accuracy under dynamic sensor networks, highlighting its practical utility for adaptive traffic management. The work advances inductive spatio-temporal modeling in dynamic sensing contexts and suggests future directions in uncertainty estimation, multi-source data fusion, and multivariate traffic data integration.

Abstract

Given a partially observed road network, how can we predict the traffic state of unobserved locations? While deep learning approaches show exceptional performance in traffic prediction, most assume sensors at all locations of interest, which is impractical due to financial constraints. Furthermore, these methods typically require costly retraining when sensor configurations change. We propose MoGERNN, an inductive spatio-temporal graph representation model, to address these challenges. Inspired by the Mixture of Experts approach in Large Language Models, we introduce a Mixture of Graph Expert (MoGE) block to model complex spatial dependencies through multiple graph message aggregators and a sparse gating network. This block estimates initial states for unobserved locations, which are then processed by a GRU-based Encoder-Decoder that integrates a graph message aggregator to capture spatio-temporal dependencies and predict future states. Experiments on two real-world datasets show MoGERNN consistently outperforms baseline methods for both observed and unobserved locations. MoGERNN can accurately predict congestion evolution even in areas without sensors, offering valuable information for traffic management. Moreover, MoGERNN is adaptable to dynamic sensing networks, maintaining competitive performance even compared to its retrained counterpart. Tests with different numbers of available sensors confirm its consistent superiority, and ablation studies validate the effectiveness of its key modules.
Paper Structure (17 sections, 16 equations, 5 figures, 5 tables, 1 algorithm)

This paper contains 17 sections, 16 equations, 5 figures, 5 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Literature Review
  3. Spatial Extrapolation
  4. Temporal Extrapolation
  5. Summary
  6. Problem Formulation
  7. Methodology
  8. Mixture of Graph Experts For Unobserved Node Embedding
  9. Graph Encoder-Decoder For Spatio-Temporal Extrapolation
  10. Model Training and Testing
  11. Numerical Experiments
  12. Experimental Setup
  13. Prediction Performance on Unobserved and Observed Locations (Q1)
  14. Adaptability of model to dynamic sensing networks (Q2)
  15. Impact of the VS-to-AAS Ratio (Q3)
  16. ...and 2 more sections

Figures (5)

  • Figure 1: An illustration of forecasting unobserved nodes under a dynamic sensing network: (a) an example of sensor deployment on the map, with some interested but not observed locations; (b) graph typologies for dynamic sensor configuration.
  • Figure 2: The overview of the proposed MoGERNN.
  • Figure 3: Map presentation of prediction performance in an evening-peak time point (2012-05-24 17:30) of METR-LA.
  • Figure 4: Predicting results for METR-LA. (a) show the results of unobserved locations, including three virtual sensors (Sensor #193, #98, #70). (b) show the results of observed locations, including one real sensor (Sensor #68).
  • Figure 5: Model performance under different ratios of VS to AAS. The first row of x-ticks labels indicates the number of VS, while the second row shows the number of AAS.