This repo is cloned from the repo chaosido/FACT-course. To see the changelog used for this project, see the original repo.

✨ TGNExplainer reproduction ✨

This repo is ment to ease the reproduction of results of TGNNExplainer. In addition this repository adds two more open-source datasets for training on the proposed methodology.

Download wikipedia and reddit datasets

Download from http://snap.stanford.edu/jodie/wikipedia.csv and http://snap.stanford.edu/jodie/reddit.csv and http://snap.stanford.edu/jodie/mooc.csv https://snap.stanford.edu/data/soc-redditHyperlinks-body.tsv and put them into ~/workspace/TGNNEXPLAINER-PUBLIC/tgnnexplainer/xgraph/dataset/data

The reddit_hyperlinks dataset should be converted to .csv format and preprocessed. Preprocessing reddit_hyperlinks can be done here.

setting up training evironment

conda env create -f conda_fact.yml

This environment is incompatible with Tick.

Preprocess real-world datasets

cd  ~/workspace/TGNNEXPLAINER-PUBLIC/tgnnexplainer/xgraph/models/ext/tgat
python process.py -d wikipedia
python process.py -d reddit
python process.py -d mooc
python process py -d reddit_hyperlinks

Generate simulate dataset

cd  ~/workspace/TGNNEXPLAINER-PUBLIC/tgnnexplainer/xgraph/dataset
python generate_simulate_dataset.py -d simulate_v1
python generate_simulate_dataset.py -d simulate_v2

This step generates the simulate datasets with Tick. note that the Tick module is depricated for Py>3.7. It is advised to create a new conda environment with py=3.7. to install the Tick module and generate the syntetic datasets.

Generate explain indexs

cd  ~/workspace/TGNNEXPLAINER-PUBLIC/tgnnexplainer/xgraph/dataset
python tg_dataset.py -d reddit -c index

This step creates a test set for the explainers. it randomly selects 500 indexes from the full test set.

Train tgat/tgn model

tgat:

cd  ~/workspace/TGNNEXPLAINER-PUBLIC/tgnnexplainer/xgraph/models/ext/tgat
./train.sh
./cpckpt.sh

tgn:

cd  ~/workspace/TGNNEXPLAINER-PUBLIC/tgnnexplainer/xgraph/models/ext/tgn
./train.sh
./cpckpt.sh

The cpckpt.sh ensures that the saved TGAT model is findable during the explainer training. make sure cpckpt.sh is ran for for each model dataset combination

Run our explainer and other baselines

cd  ~/workspace/TGNNEXPLAINER-PUBLIC/benchmarks/xgraph
./run_explainers_model_dataset.sh

In the benchmars directory a shell script exists for training all 4 explainers on a (dataset,model) combination.

dataset= reddit, Wikipedia, simulate_v1, simulate_v2, mooc, reddit_hyperlinks. model= TGAT,TGN