PASTEL

Code for "Position-aware Structure Learning for Graph Topology-imbalance by Relieving Under-reaching and Over-squashing" (accepted by CIKM 2022).

Overview

• main.py: getting started.
• src/model.py: model configurations.
• src/model_handler.py: overall framework of PASTEL.
• src/models/graph_clf.py: configurations for GNN backbones and the graph structure learners.
• src/layers/backbones.py: the basic GNNs we used as the backbone, including GCN, GAT, APPNP and GraphSAGE.
• src/layers/graph_learner.py: graph learner of the input graph.
• src/utils/*.py: utils for data preprocessing, loading and other functions.

Requirements

The implementation of PASTEL is tested under Python 3.8.12, with the following packages installed:

• dgl==0.8.0.post1
• dgl_cu111==0.7.2
• networkx==2.6.3
• numpy==1.20.3
• PyYAML==6.0
• scikit_learn==1.1.1
• scipy==1.7.1
• torch==1.9.1+cu111
• torch_geometric==2.0.1

You can install the above packages by pip:

• pip install -r requirements.txt

Or by conda:

• conda install --yes --file requirements.txt

Datasets

All the datasets (i.e., Cora, Citeseer, Photo, Actor) are provided by pytorch_geometric.

Run the codes

Train and evaluate the model:

• python main.py -config config/<dataset>.yml

We train the PASTEL with GNN backbones, and report the training loss, validation loss, training accuracy, validation accuracy and the test accuracy. For instance:

• python main.py -config config/cora.yml

Train Epoch [ 50 / 10000] | Loss: 5.94542 | NLOSS = -5.94542 | ACC = 0.60714
  Val Epoch [ 50 / 10000] | Loss: 4.11519 | NLOSS = -4.11519 | ACC = 0.63333
Train Epoch [100 / 10000] | Loss: 98.20475 | NLOSS = -98.20475 | ACC = 0.80714
  Val Epoch [100 / 10000] | Loss: 36.39100 | NLOSS = -36.39100 | ACC = 0.82381
Train Epoch [150 / 10000] | Loss: 9.87782 | NLOSS = -9.87782 | ACC = 0.89286
  Val Epoch [150 / 10000] | Loss: 5.40413 | NLOSS = -5.40413 | ACC = 0.80952
Train Epoch [200 / 10000] | Loss: 3.95029 | NLOSS = -3.95029 | ACC = 0.83571
  Val Epoch [200 / 10000] | Loss: 2.70414 | NLOSS = -2.70414 | ACC = 0.82857
Train Epoch [250 / 10000] | Loss: 2.22446 | NLOSS = -2.22446 | ACC = 0.85000
  Val Epoch [250 / 10000] | Loss: 1.83093 | NLOSS = -1.83093 | ACC = 0.84286
Train Epoch [300 / 10000] | Loss: 1.50548 | NLOSS = -1.50548 | ACC = 0.89286
  Val Epoch [300 / 10000] | Loss: 1.61285 | NLOSS = -1.61285 | ACC = 0.83333
......
Finished Training. Training time: 1930.99
********************** MODEL SUMMARY **********************
Best epoch = 707 | NLOSS = -1.11036 | ACC = 0.85238
******************** END MODEL SUMMARY ********************
[test] | ACC: 0.82619 | W-F1: 0.82415 | M-F1: 0.81198
Finished Testing. Testing time: 0.12