Mis on lingi prognoosimine ja milliseid arhitektuure kasutatakse?

Link prediction is a task in graph analysis where the goal is to predict the probability of an edge existing between two nodes, based on the graph's structure and the properties of the nodes.

Main architectures and approaches for link prediction:

• Feature-based methods: utilize characteristics of nodes and their neighbors, such as common neighbors, Jaccard coefficient.

• Graph Neural Networks (GNN): train representations of nodes considering their environment. Popular architectures:

  • Graph Convolutional Networks (GCN)
  • Graph Attention Networks (GAT)
  • GraphSAGE

• Embedding-based models: nodes are represented in a vector space, and the connection is predicted through operations on vectors (e.g., dot product).

• Probabilistic graph models: use probabilistic models to evaluate the links.

Example with GNN for link prediction: first, train a GNN to obtain node embeddings, then compute the probability of an edge between two nodes using a similarity function (e.g., dot product), and apply a sigmoid to get the probability.