Graph machine learning with missing node features

Graph Neural Network (GNN) models typically assume a full feature vector for each node. The two inputs to this model are the (normalised) adjacency matrix A encoding the graph structure and the feature matrix X containing as rows the node feature vectors and outputs the node embeddings Z. Each layer of GCN performs node-wise feature transformation (parametrised by the learnable matrices W₁ and W₂) and then propagates the transformed feature vectors to the neighboring nodes. Importantly, GCN assumes that all the entries in X are observed. In real-world scenarios, we often see situations where some node features are missing (Fig 1).