About The Project

Code repo from my master thesis from 2022, in which we proposed four novel road segment representation learning models. The datasets and model states were removed because of storage limits. Furthermore, the implementation of the toast model was removed because of copyright conflicts.

Abstract:

Road networks, as the core component of urban transportation, are naturally highly dynamic, considering traffic patterns and movement behavior induced by traversing vehicles. Deriving robust road network representations that capture the underlying dynamics and spatial properties of the road network can directly improve the effectiveness of downstream tasks, such as travel time or destination predictions. Although, various approaches in the field of representation learning for road networks have been recently proposed, they focus on spatial properties without capturing the dynamic characteristics. In this work, we first propose the Graph Trajectory Convolution (GTC) model, which incorporates traffic flow information induced by real trajectory data into the aggregation process. We transfer the idea of modeling traffic flows directly in the learning process to the DeepWalk architecture and subsequently propose the Trajectory Smoothed DeepWalk (TSD) model. We further combine the mentioned models with a transformer architecture, resulting in the Graph Trajectory Network (GTN) model, which additively optimizes the representations to capture multidimensional trajectory traveling behavior and structural properties of the road network. We provide extensive experiments including comprehensive benchmarking, a parameter study, trajectory feature analysis, and an embedding analysis, showing the superior performance of our GTN model across five different intelligent transportation systems (ITS) downstream tasks and three different real-world road networks. The GTN model achieved a 57% higher generalization performance than the best performing baseline model. Furthermore, we propose another novel model architecture, namely Temporal Graph Trajectory Convolution (T-GTC), which incorporates the temporal dependence of road networks into the road segment representations. This model learns the temporal de- pendence of the road network by considering dynamic properties, such as driving speed measurements. The T-GTC model achieved a 30.4% higher generalization performance over the downstream tasks and a 7.3% better mean absolute error (MAE) on travel time prediction than the best performing non-temporal baseline, showing that the incorporation of temporal dynamics significantly boosts the performance on temporal dependent tasks.

Implemented Road Network Representation Learning Baselines:

All models are written in Python using Pytorch (except for rfn, which uses mxnet)

Getting Started

WIP

Prerequisites

To install the required packages we recommend using Conda. Our used environment can be easily installed with conda.

Installation

  1. Clone the repo
    git clone https://github.com/...
  2. Install conda environment
    conda config --env --set channel_priority strict
    conda env create -f environment.yml
  3. Activate the environment
    conda activate road-emb
  4. Install Fast Map Matching in the environment (Do the steps while in environment) and for MacOS do also this

Tests

WIP

Reproduction

WIP

Contact

  1. Paul Heinemeyer - paul_heinemeyer@outlook.de