GAIN

Overview

This repository contains the codes of a novel approach, GAIN, to aggregating knowledge for learning representations by graph convolutional neural networks. The GAIN architecture is developed to address multi-class road type classification problem inspired by GraphSAGE. Road network graph datasets are generated from OpenStreeMap (OSMnx) and preprocessed according to the corresponding settings. Representation learning improves by application of a search mechanism in the local and the global neighborhood of a graph node. Anyone interested in using GAIN architecture, please cite the following paper:

@article {gharaee2021pr} {
  author = {Gharaee, Zahra and Kowshik, Shreyas and Stromann, Oliver and Felsberg, Michael},
    title = {Graph representation learning for road type classification},
    booktitle = {Pattern Recognition},
    year = {2021}
    page = {}
    volume = {120}
    DOI = {https://doi.org/10.1016/j.patcog.2021.108174}
  }
}

Requirement

Use the packages mentioned in the requirements.txt file to generate road network graphs and to run experiments.

Road network graph generation

Run roadnetwork_graphs.py scripts in codes folder in order to generate transductive road network graphs of Linköping city and inductive road network graphs of Sweden country extracted from OpenStreetMap (OSMnx). Running roadnetwork_graphs.py also generates id-map, class-map, raw features/attributes and the pairs of topological neighbors. A set of transductive and inductive road network graphs of Linköping city and Sweden country are available in graph_data_GainRepo folder.

Image below shows road network graph of the Linköping city representing our transductive data set. Road-type class labels of the original graph are described as following and its line graph representation is overlaid in black:

Class1: red
Class2: orange
Class3: yellow
Class4: skyblue
Class5: lime


Road network graph of Linköping city area.

Road types classification

run_supervised.py and run_unsupervised.py python scripts contans the main codes and the configurations of hyperparameters to run experiments for supervised and unsupervised settings, respectively.

Results

osm_eval.py python script evaluates the representation vectors generated by GAIN for road type classification. Running this script shows the performance results of applying random-baseline, raw-features, and representation vectors of a pre-trained GAIN model, available in logs_GainRepo folder, to classify road networks of transductive and inductive test datasets in both supervised and unsupervised settings.

zahrag / GAIN