BSG

BSG introduces three loss functions.

The details of each loss function are included in the model.py line 457 - 467.

######### BSG Part!
            z_unmask = self.encoder(x, masked_edges)
            loss += alpha * F.mse_loss(z, z_unmask)

            aggregated_message = self.mean_aggregator(z, edge_index=remaining_edges)
            loss += beta * F.mse_loss(z, aggregated_message)

            similarity_matrix_edge = self.cosine(z, aggregated_message)
            loss += gamma * torch.mean(torch.relu(similarity_matrix_edge - margin))
#### 

We utilized the implementation of the baseline MaskGAE as our backbone structure.

Environment

Higher versions should also be available.

• numpy==1.26.3
• torch==2.1.2+cu121
• torch-cluster==1.6.3
• torch_geometric==2.4.0
• torch-scatter==2.1.2
• torch-sparse==0.6.18
• scipy==1.11.4
• texttable==1.7.0

Newly added hyperparameters

• alpha = $\lambda_2$
• beta = $\lambda_1$
• gamma = $\lambda_3$
• margin = margin $m$

Reproduction

Link Prediction

python train_link.py --dataset <dataset_name>

<dataset_name> = [Cora, Citeseer, Pubmed, Photo, Computers]

The image is the hyperparamter configuration, which corresponds to Table 6 in our paper.

Node Classification

example with Cora

python train_node.py --dataset Cora --alpha 0.0002 --beta 0.001 --gamma 0.0009  --margin -0.2