This library contains a reproducted implementation of the Relational Fusion Network (RFN)
introduced in the SIGSPATIAL'19 paper Graph Convolutional Networks for Road Networks
by Tobias Skovgaard Jepsen, Christian S. Jensen, and Thomas Dyhre Nielsen.
Paper can be found in here
and code can be found in here .
Most of the code is from the original code.
Custom added codes are in custom folder.
Road network data for 12 Korean cities are in data.zip.
First unzip it.
Pre-trained network parameter data is in model_data folder.
See Train.ipynb for training, Test.ipynb for testing.
You need to install osmnx, MXNet, dgl, PyTorch. MXNet and dgl should be gpu-version.
For node attribute, in/out degree were used. For edge attribute, road category, length were used. For between-edge attribute, turning angle was used.
Original paper used historical speed data in single city and
split them into train/test data.
However, I could not obtain such date-by-date data.
Instead, I used osmnx default provided driving speed data.
However they gave speed values with limited types.
(e.g. only 7 different values in Suwon)
If you can access more accurate speed data, it is recommended to use it instead.
MSE loss for test set cities are written down in below table.
| Daegu | Suwon | Ulsan | Yongin | average | |
|---|---|---|---|---|---|
| RFN_Int_Att | 0.01380 | 0.00647 | 0.01754 | 0.00821 | 0.01150 |
| RFN_Int_Non | 0.01725 | 0.01252 | 0.01902 | 0.02812 | 0.01923 |
| RFN_Add_Att | 0.01598 | 0.01097 | 0.01968 | 0.02828 | 0.01873 |
| RFN_Add_Non | 0.01582 | 0.01013 | 0.01954 | 0.02983 | 0.01883 |
| GAT | 0.04189 | 0.02764 | 0.03788 | 0.05231 | 0.03993 |
| GCN | 0.04681 | 0.03600 | 0.04592 | 0.06588 | 0.04865 |
| GraphSAGE | 0.04487 | 0.03236 | 0.03112 | 0.04475 | 0.03827 |
| MLP | 0.02547 | 0.01225 | 0.02236 | 0.01405 | 0.01853 |
RFN gave better result than baseline algorithms. Especially RFN with intentional fusion and attentional aggregation gave the best result. Although limited data was used, it was possible to verify the effectiveness of RFN over traditional graph convolution networks as the paper says.