Graph Classification Exercise

• ogbg-molhiv dataset from Open Graph Benchmark:
• Predict whether a molecule (graph) inhibits HIV replication
• 41k molecules from MoleculeNet
• Average of 26 nodes and 28 edges per graph
• Natural features for both atoms (nodes) and bonds (edges):
• Categorical in nature and the features are therefore integers that indicate the category mapping for each column
• List of features is here, but include things like atomic number, bond type, …, etc.
• OGB provides classes for converting these categorical/integer features into embeddings: AtomEncoder and BondEncoder
• Split into train/val/test based on molecular structure
  • To evaluate generalization, the test nodes have structural differences from those in the training set