baldassarrefe/graph-network-explainability

Explainability techniques for Graph Networks, applied to a synthetic dataset and an organic chemistry task. Code for the workshop paper "Explainability Techniques for Graph Convolutional Networks" (ICML19)

Explainability Techniques for Graph Convolutional Networks

Code and notebooks for the paper "Explainability Techniques for Graph Convolutional Networks" accepted at the ICML 2019 Workshop "Learning and Reasoning with Graph-Structured Data".

Overview

A Graph Network trained to predict the solubility of organic molecules is applied to sucrose, the prediction is explained using Layer-wise Relevance Propagation that assigns positive and negative relevance to the nodes and edges of the molecular graph:

The predicted solubility can be broken down to the individual features of the atoms and their bonds:

Code structure

• src, config, data contain code, configuration files and data for the experiments
  • infection, solubility contain the code for the two experiments in the paper
  • torchgraphs contain the core graph network library
  • guidedbackrprop, relevance contain the code to run Guided Backpropagation and Layer-wise Relevance Propagation on top of PyTorch's autograd
• notebooks, models contain a visualization of the datasets, the trained models and the results of our experiments
• test contains unit tests for the torchgraphs module (core GN library)
• conda.yaml contains the conda environment for the project

Setup

The project is build on top of Python 3.7, PyTorch 1.1+, torchgraphs 0.0.1 and many other open source projects.

A Conda environment for the project can be installed as:

conda env create -n gn-exp -f conda.yaml
conda activate gn-exp
python setup.py develop
pytest

Training

Detailed instructions for data processing, training and hyperparameter search can be found in the respective subfolders:

• Infection: infection/notes.md
• Solubility: solubility/notes.md

Experimental results

The results of our experiments are visualized through the notebooks in notebooks:

conda activate gn-exp
cd notebooks
jupyter lab 

graph-network-explainability's Issues

Question on how Message Passing is performed on torchgraphs

I am currently trying to implement LRP on my own graph neural network, focused on discovering relevances at the node feature level (on my own implementation) where node features are roughly updated as:

H_next = (A.dot(H)).dot(W)

I see however, that the torchgraph implementation, and thus the LRP step in your research solely multiplies the node features with the weights.

Thus I would like to better understand how the node adjacencies are taken into consideration for the computations?

Thank you, and thank you for the great research and examples!

Unable to setup the conda env & not able to install torchgraph

Hi,

The conda env installation/ creation is failing:
conda env create -n gn-exp -f conda.yaml
Collecting package metadata (repodata.json): done
Solving environment: failed

ResolvePackageNotFound:

• fontconfig==2.13.0=h9420a91_0
• zstd==1.3.7=h0b5b093_0
• glib==2.56.2=hd408876_0
• libgcc-ng==8.2.0=hdf63c60_1
• cffi==1.12.2=py37h2e261b9_1
• libtiff==4.0.10=h2733197_2
• kiwisolver==1.0.1=py37hf484d3e_0
• cudatoolkit==10.0.130=0
• pyqt==5.9.2=py37h05f1152_2
• pillow==5.4.1=py37h34e0f95_0
• icu==58.2=h9c2bf20_1
• libedit==3.1.20181209=hc058e9b_0
• xz==5.2.4=h14c3975_4
• libxcb==1.13=h1bed415_1
• libffi==3.2.1=hd88cf55_4
• mkl_random==1.0.2=py37hd81dba3_0
• python==3.7.3=h0371630_0
• pixman==0.38.0=h7b6447c_0
• mkl_fft==1.0.12=py37ha843d7b_0
• gmp==6.1.2=h6c8ec71_1
• libgfortran-ng==7.3.0=hdf63c60_0
• tornado==6.0.2=py37h7b6447c_0
• libuuid==1.0.3=h1bed415_2
• pcre==8.43=he6710b0_0
• py-boost==1.67.0=py37h04863e7_4
• yaml==0.1.7=had09818_2
• gst-plugins-base==1.14.0=hbbd80ab_1
• libsodium==1.0.16=h1bed415_0
• readline==7.0=h7b6447c_5
• pytorch==1.1.0=py3.7_cuda10.0.130_cudnn7.5.1_0
• freetype==2.9.1=h8a8886c_1
• sip==4.19.8=py37hf484d3e_0
• rdkit==2019.03.1.0=py37hc20afe1_1
• openssl==1.1.1c=h7b6447c_1
• numpy-base==1.16.4=py37hde5b4d6_0
• scipy==1.2.1=py37h7c811a0_0
• ncurses==6.1=he6710b0_1
• jpeg==9b=h024ee3a_2
• mistune==0.8.4=py37h7b6447c_0
• qt==5.9.7=h5867ecd_1
• dbus==1.13.6=h746ee38_0
• libpng==1.6.36=hbc83047_0
• expat==2.2.6=he6710b0_0
• sqlite==3.27.2=h7b6447c_0
• libboost==1.67.0=h46d08c1_4
• zlib==1.2.11=h7b6447c_3
• tk==8.6.8=hbc83047_0
• libstdcxx-ng==8.2.0=hdf63c60_1
• pandas==0.24.2=py37he6710b0_0
• libxml2==2.9.9=he19cac6_0
• bzip2==1.0.6=h14c3975_5
• matplotlib==3.0.3=py37h5429711_0
• cairo==1.14.12=h8948797_3
• gstreamer==1.14.0=hb453b48_1
• pyrsistent==0.14.11=py37h7b6447c_0
• pyyaml==5.1=py37h7b6447c_0
• markupsafe==1.1.1=py37h7b6447c_0
• cudnn==7.3.1=cuda10.0_0
• pyzmq==18.0.0=py37he6710b0_0
• ninja==1.9.0=py37hfd86e86_0
• zeromq==4.3.1=he6710b0_3
• scikit-learn==0.20.3=py37hd81dba3_0

and so, tried to install independently torchgraph:
$ pip install --user https://github.com/baldassarreFe/torchgraphs

Collecting https://github.com/baldassarreFe/torchgraphs
Downloading https://github.com/baldassarreFe/torchgraphs
- 121 kB 1.2 MB/s
ERROR: Cannot unpack file /private/var/folders/mj/fk1t3xm53gn_sd13cv2dq22c0000gp/T/pip-unpack-wt1i2ex1/torchgraphs (downloaded from /private/var/folders/mj/fk1t3xm53gn_sd13cv2dq22c0000gp/T/pip-req-build-xyu4u4km, content-type: text/html; charset=utf-8); cannot detect archive format
ERROR: Cannot determine archive format of /private/var/folders/mj/fk1t3xm53gn_sd13cv2dq22c0000gp/T/pip-req-build-xyu4u4km

problems I met when I was running the code for solubility

Hi! I met some problems when I was running the predict part of code for solubility. I'm wondering how can I find the datasets for this part, because it seems that some ".yaml" files are absent?