Questions regarding Graph Autoencoder and reconstructing input features + edge indices

[sivaramkrishnan101]

Hello everyone,

Problem - I'm currently working on implementing a Graph Autoencoder using a custom dataset.

Scenario - The dataset consists of 20 graphs, each with 5 nodes and 2 features. These graphs are fully connected, and I have generated node embeddings using the SAGECONV method. z = SAGECONV(2, 2) -> SAGECONV(2, 1)

Questions - How can I define a decoder architecture to

• Reconstruct the input matrix (I am thinking of using the FCN layers for this)?
• Reconstruct the adjacency matrix.

I hope to get some ideas on this implementation as I have not seen a GAE defined to reconstruct the input features so far. Thank you :)

[EdisonLeeeee]

You might find the following repositories helpful:

• Reconstruct features: https://github.com/THUDM/GraphMAE
• Reconstruct structure: https://github.com/EdisonLeeeee/MaskGAE

[sivaramkrishnan101]

Hi Edison,

I appreciate your response. This is exactly what I was looking for. Thank you very much!

[sujoyrc]

what if I want to recover both with a weighted loss between the two components. My node features are text and I can embed the text via a SentenceTransformer model into an embedding. every node has only 1 text