Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

GNN-based Passenger Request Prediction

Aqsa Ashraf Makhdomi, Iqra Altaf Gillani

TL;DR

This work tackles OD prediction for ride-hailing by modeling the city as a sequence of graphs and applying a Graph Attention Network with specialized spatial, temporal, and transferring attention layers. It introduces a pre-weighted aggregator to bias neighbor influence using OD flows and distances, and a non-linear temporal channel to capture context-aware events and travel patterns, complemented by a transferring layer for OD pair estimation. The approach is validated on NYC and Washington DC datasets, with extensive ablations showing the model outperforms baselines across OD and demand tasks, and revealing the importance of grid size and horizon length for capturing dependencies. Practically, the framework enables more accurate and context-aware predictions, supporting better driver allocation, reduced wait times, and potential environmental and economic benefits for ride-hailing platforms.

Abstract

Passenger request prediction is essential for operations planning, control, and management in ride-sharing platforms. While the demand prediction problem has been studied extensively, the Origin-Destination (OD) flow prediction of passengers has received less attention from the research community. This paper develops a Graph Neural Network framework along with the Attention Mechanism to predict the OD flow of passengers. The proposed framework exploits various linear and non-linear dependencies that arise among requests originating from different locations and captures the repetition pattern and the contextual data of that place. Moreover, the optimal size of the grid cell that covers the road network and preserves the complexity and accuracy of the model is determined. Extensive simulations are conducted to examine the characteristics of our proposed approach and its various components. The results show the superior performance of our proposed model compared to the existing baselines.

GNN-based Passenger Request Prediction

TL;DR

This work tackles OD prediction for ride-hailing by modeling the city as a sequence of graphs and applying a Graph Attention Network with specialized spatial, temporal, and transferring attention layers. It introduces a pre-weighted aggregator to bias neighbor influence using OD flows and distances, and a non-linear temporal channel to capture context-aware events and travel patterns, complemented by a transferring layer for OD pair estimation. The approach is validated on NYC and Washington DC datasets, with extensive ablations showing the model outperforms baselines across OD and demand tasks, and revealing the importance of grid size and horizon length for capturing dependencies. Practically, the framework enables more accurate and context-aware predictions, supporting better driver allocation, reduced wait times, and potential environmental and economic benefits for ride-hailing platforms.

Abstract

Passenger request prediction is essential for operations planning, control, and management in ride-sharing platforms. While the demand prediction problem has been studied extensively, the Origin-Destination (OD) flow prediction of passengers has received less attention from the research community. This paper develops a Graph Neural Network framework along with the Attention Mechanism to predict the OD flow of passengers. The proposed framework exploits various linear and non-linear dependencies that arise among requests originating from different locations and captures the repetition pattern and the contextual data of that place. Moreover, the optimal size of the grid cell that covers the road network and preserves the complexity and accuracy of the model is determined. Extensive simulations are conducted to examine the characteristics of our proposed approach and its various components. The results show the superior performance of our proposed model compared to the existing baselines.
Paper Structure (20 sections, 14 equations, 13 figures, 8 tables)

This paper contains 20 sections, 14 equations, 13 figures, 8 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Preliminaries
  4. System model
  5. Spatial Attention Layer
  6. Pre-weighted aggregator
  7. Temporal Attention Layer
  8. Context-aware data
  9. Travelling Behavior
  10. Transferring Attention Layer
  11. Prototype Implementation
  12. Grid cell length
  13. Non-linearities
  14. Parameter setting verification with different models and data distributions
  15. Prediction accuracy with different time slot granularities
  16. ...and 5 more sections

Figures (13)

  • Figure 1: Road network represented in the form of a grid
  • Figure 2: OD Matrix of the grid
  • Figure 3: Framework of our proposed model
  • Figure 4: Forward and backward neighbors calculated from an instance of graph $G$
  • Figure 5: Graph Attention Network for calculating affinity between neighboring nodes
  • ...and 8 more figures