Improving Training (Ideas)

[faseelmo]

• [x] Use Larger Network
  i.e Denser and Wider Layers.
• [x] Using ResNet inspired network for a deeper GCN.
• [ ] Using Concatenation ?
• [ ] Adding batch normalization after Conv layers.
• [ ] Normalizing target based on their Mapping.
  i.e normalising mappings of every unique graph between 0-1 with its own max and min mapping variation.
• [ ] Think about implementing ranking based loss function?
• [x] Use directed graph (the final graph that is converted to tensor)
• [x] Implement heterogeneous graph learning. heterogeneous
• [ ] Implement Directed Graph Learning directed graph.
• [ ] Giving more weight to link -> router edges?
• [ ] Loading Data sequentially. Use the 100 mapping of a unique task as a batch.
• [x] Using LSTM for sequential data (flow of data). For example, passing node embedding to a LSTM or replacing the MLP with LSTM.
• [ ] Adding Positional Encoding information to Nodes ? i.e Degree, Node2Vec or DeepWalk or GraphTransformerNetwork, Global structure using Laplacian Eigenvector.