Module graphmae 

GraphMAE: Masked Autoencoders for Graphs

Self-supervised graph learning via masked feature reconstruction. Traditional supervised graph learning requires expensive node/edge labels that are scarce in real-world graphs. GraphMAE learns representations by masking and reconstructing node features, requiring zero labels. The learned embeddings transfer well to downstream tasks (classification, link prediction, clustering) because the model must capture structural and semantic graph properties to reconstruct masked features from their neighborhood context.

Pipeline: Mask -> GAT Encode -> Re-mask latent -> Decode masked only -> SCE loss.

Reference: Hou et al., “GraphMAE: Self-Supervised Masked Graph Autoencoders”, KDD 2022.

Structs

FeatureMasking

Feature masking strategies for GraphMAE.

GATEncoder

Multi-layer GAT encoder for GraphMAE.

GraphData

Sparse graph representation.

GraphMAE

GraphMAE self-supervised model.

GraphMAEConfig

Configuration for a GraphMAE model.

GraphMAEDecoder

Decoder that reconstructs only masked node features (key efficiency gain).

MaskResult

Result of masking node features.

Enums

LossFn

Loss function variant for reconstruction.

Functions

mse_loss

Mean Squared Error across masked node reconstructions.

sce_loss

Scaled Cosine Error: mean((1 - cos_sim(pred, target))^gamma) over masked nodes.