Link Prediction Models

Best for: Predicting relationships

How it works

Predicts whether an edge should exist between nodes $i$ and $j$. Heuristic scores use neighbourhood overlap such as Adamic-Adar $\sum_{k\in N(i)\cap N(j)}1/\log d_k$ or Jaccard similarity of neighbour sets. Embedding-based approaches score a pair by $\hat y_{ij}=\sigma\bigl(f(z_i,z_j)\bigr)$, e.g. the dot product of node2vec or GNN embeddings, while GNN variants pool over $(z_i,z_j)$ and classify. Trained on observed edges versus sampled negative edges.

Common fields

Knowledge graphs · social networks · drug discovery