A framework for graph learning
If we take a holistic view of the approaches that are followed for learning inference models on graphs, we’ll notice a pattern. Every solution can be divided into three distinct steps:
- The first step involves coming up with a mechanism to find a local subgraph, given a node in the graph. This term needs to be defined here. For example, the graph containing all the nodes that are directly connected to an edge of the concerned node can be a local subgraph. Another example can be the set of nodes that have a first or second-degree connection to the concerned node. This local subgraph is often called the receptive field of the concerned node in academic literature.
- The second step involves a mechanism that takes input from the concerned node and its receptive field and outputs the node embedding. The node embedding is simply a vector of real values of a certain dimension. It’s important to have a similarity metric defined in this...
