Implementation for paper Link Prediction with Persistent Homology: An Interactive View (ICML 2021, TLC-GNN) and for paper Neural Approximation of Graph Topological Features (NeurIPS 2022, PDGNN)
Python version is 3.7, and the versions of needed packages are listed in requirements.txt
We also provide a packed environment, which is available at:
you can activate the environment with:
tar -xzf curvGN.tar.gz -C curvGN
source curvGN/bin/activate
python pipelines.py
to run the experiments for PubMed, Photo and Computers datasets, the results will be stored in ./scores.
If you want to run experiments for PPI datasets, you can comment out line 56 in pipelines.py.
For Cora and Citeseer, you can set the dropout in ./baselines/TLC-GNN to 0.8, the results can be a little higher
cd ./sg2dgm
python setup_PI.py build_ext --inplace
to setup ./sg2dgm/persistenceImager.pyx
If the command does not work, a substitute solution is to copy the code in ./sg2dgm/persistenceImager.pyx to a new file named ./sg2dgm/persistenceImager.py, this might also work.
A small note: PDGNN has nothing to do with knowledge distillation, actually.
First generate the training data for node-centered vicinity graphs:
python ./Knowledge_Distillation/data_utils_NC.py
You can also generate edge-centered vicinity graphs / total graphs using
python ./Knowledge_Distillation/data_utils_LP.py
python ./Knowledge_Distillation/data_utils_GC.py
Then train PDGNN, notice that the saved data dir need to be revised in lines15-26 in ./Knowledge_Distillation/train_Teacher_Model.py.
Notice that in the code, PDGNN is actually in ./Knowledge_Distillation/gat_conv.py rather than ./Knowledge_Distillation/PD_conv.py
python ./Knowledge_Distillation/train_Teacher_Model.py
You can also train PDGNN on TUdatasets, ZINC, and OGBG-HIV
python ./Knowledge_Distillation/train_Teacher_Model_GC.py
Afther training PDGNN, your can run downstream node classification/link prediction tasks with
python pipelines_GIN.py
python pipelines_LP_GIN.py
