GraphReach

GraphReach is a position-aware inductive GNN that captures the global positions of nodes through reachability estimations with respect to a set of anchor nodes. The anchors are strategically selected so that reachability estimations across all the nodes are maximized. This strategy also makes the model more robust to adversarial attacks.

Please cite our paper, if you use our source code.

• "GraphReach : Position-Aware Graph Neural Network using Reachability Estimations, IJCAI'21"

Code Structure

There are three code folders. Two of them are for GraphReach with the following functionalities :

• Link Prediction and Pairwise Node Classification (src_GraphReach)
• Node Classification (src_GraphReach_NC)

One for P-GNN with adversarial attack functionality (pgnn_adversarial)

The P-GNN framework was taken from [https://github.com/JiaxuanYou/P-GNN] and has other benchmarks implemented as well. The framework code is added as the adversarial module is additional to the existing prior work.

Experimental Setup

• Python 3.7.6
• Pytorch 1.4.0
• NetworkX 2.3
• Cuda 10.0

Libraries Required

• Pytorch
• PyTorch Geometric
• torch-scatter
• torch-sparse
• torch-cluster
• torch-spline-conv (optional)
• networkx
• tensorboardx

USAGE

The models can be run using the shell file in respective folders. USAGE: ./cmds.sh

Arguments:

Note: Arguments related to random walks correspond to GraphReach only

• --model: The model to be executed. Possible values: GraphReach, PGNN, GCN, MLP, SAGE, GAT, GIN
• --dataset: Dataset to train the model. Possible values: grid, communities, ppi, email, completeEmail, protein, All
• --task: Task to perform, link prediction (LP) or pairwise node classification (PNC). Possible values: link (for LP), link_pair (for PNC)
• --remove_link_ratio: Fraction of data to be retained as validation and test set (Validation and test are split into half)
• --layer_num: Number of layers in model
• --epoch_num : Number of epochs
• --repeat_num: Number of iterations of experiment
• --num_walk: Number of random walks
• --walk_length: Length of the random walk
• --unweightedRandomWalk: For unweighted random walk
• --weightedRandomWalk: For weighted random walk
• --edgelabel: To bias random walk. Used only for pairwise node classification task. Not to be used in adversarial setting
• --Num_Anchors: Number of anchors. Possible values: 3, 5, 8, logn, 2logn, 5logn, logn2 (logn2 for (log(n))^2 anchors)
• --select_anchors: Anchor Selection Strategy (For GraphReach only). Possible values: DiversifiedRandomK, DiversifiedTopK, topK, random
• --sampleXwalks: Sample X% of walks from all the random walks (For diversified anchor selection)
• --sampleMbigraphs: Sample M Bipartite Graphs (For diversified anchor selection)
• --rm_feature: For inductive setting, default is tranductive
• --attention: For attention based aggregation (GraphReach only)
• --attentionAddSelf: For attention based aggregation with self feature (GraphReach only)
• --AdversarialAttack: Perform adversarial testing
• --addFedges: Add false edges to nodes involved in sampled fraction of test pairs

Testing on Unseen Graph

The model can be run on completely new unseen graph by uncommenting Line#47 in test.py. USAGE: ./test.sh

Data

Data used for the experiments can be downloaded from the following link. https://github.com/JiaxuanYou/P-GNN/tree/master/data