Link Prediction (Supervised)

This is one of the essential tasks in graph machine learning. Link prediction anticipates missing edges or prospective edges in a graph. In the case of a dynamic graph, the edges disappear and form based on the time point of the graph.

In the case of biological networks, it's difficult to identify all possible interactions using standard lab experiments like protein-protein interactions. This is why link prediction becomes crucial in this domain.

Depending upon the type of graph, link prediction can be formulated in different ways.

As binary classification

For binary classification, we need two labels, for example, positive and negative labels. We take the edges in the graph to create the instances of positive labels and randomly sample negative edges to create the instances of negative labels.

Since an edge consists of two entities (two nodes), we do an element-wise multiplication of the node embeddings that correspond to the edge to get one feature vector for every instance (positive or negative). Now, these features and labels are used to build a predictive model. Graph neural networks can be used to tackle the link prediction problem as a binary classification problem.

Here's the simplified version of the explanation above:

• Sample negative edges and positive edges.

• Take the Hadamard product of both nodes in edges to get one feature vector. ...