This repository maintains codes used for two papers [1, 2]. The scattering transform on graphs uses multi-layer and multi-scale structure based on graph wavelets and achieves robust representation of input graph signals. We specifially illustrate a graph generation model based on the graph scattering transform, which is naturally composed of an encoder and a decoder. The encoder is a Gaussianized graph scattering transform, which is robust to signal and graph manipulation. The decoder is a simple fully connected network that is adapted to specific tasks, such as link prediction, signal generation on graphs and full graph and signal generation.
This repository contains the graph scattering transform. To use it, see scat.py and the comments in that file. See example.py for the illustrative example of Cora in [1]. This repository also specifically demonstates the implementation of the experiments described in [2]. The codes can be used for implementing the following models:
- Link prediction (SCAT-S, SCAT-D)
python trainCora.py -s S [S/D] -d cora [cora/citeseer/pubmed]
- Signal generation on graph (SCAT-SW, SCAT-DW, SCAT-SN, SCAT-DN)
python trainFashion.py -s S [S/D] -g W [W/N]
- Graph and signal generation (SCAT-SW, SCAT-DW, SCAT-SN, SCAT-DN)
python -W ignore trainQM9.py -s S [S/D] -g W [W/N]
- Python >= 3.6
- Tensorflow >= 1.2.0
- RDKit
- networkx
- scipy, numpy, scikit-learn
[1] Zou, D., & Lerman, G. (2018). Graph Convolutional Neural Networks via Scattering. Applied and Computational Harmonic Analysis, in press (available online 13 June 2019), DOI: 10.1016/j.acha.2019.06.003.
[2] Zou, D., & Lerman, G. (2019). Encoding Robust Representation for Graph Generation. International Joint Conference on Neural Networks (IJCNN) 2019.
This project is licensed under the MIT License - see the LICENSE.md file for details
We borrowed data and codes from various repositories we cited in [1].