Graph Link Prediction Techniques, Self‑Developed GNN Models, and Applications in Risk Control

The article introduces graph link prediction, defining it as the task of predicting unseen or future edges in a graph and distinguishing it from node‑level and graph‑level classification.

Link prediction can be classified by graph dynamics (static vs. dynamic), node type (single vs. multi‑type, e.g., user‑item bipartite graphs), and edge relation type (single vs. multi‑relation, e.g., knowledge graphs).

Three representative application scenarios are highlighted: protein‑protein interaction discovery in biochemistry, friend recommendation in social networks, and abnormal relationship detection in risk‑control systems.

Existing techniques are surveyed:

Heuristic methods compute structural similarity using statistics such as common neighbors; they are simple but have weak generalization.

Shallow embedding methods (DeepWalk, LINE, Node2Vec) learn node vectors and compare them, capturing structure better than heuristics but often ignoring node attributes.

Graph Neural Networks (GNNs) encode node neighborhoods and predict edge scores, offering strong representation power and the ability to fuse node attributes, though many generic GNNs ignore link‑prediction‑specific cues.

To address GNN limitations, the authors propose several self‑developed encoders:

Neighborhood Attention Network (NAN) : uses multi‑head attention on each node’s neighbors and bilinear scoring of node embeddings.

Cross‑Neighborhood Attention Network : additionally attends to the neighbor set of the opposite node, capturing cross‑neighborhood interactions.

HalpNet : aggregates node and neighbor embeddings via element‑wise multiplication and hierarchical attention to form a sub‑graph representation.

Neighborhood Interaction Attention Network : computes pairwise interactions among all neighbors, applies self‑attention to select important interactions, and aggregates them for final edge scoring.

Negative sampling strategies are discussed, including global random sampling, local random sampling, and a novel adversarial negative‑sampling generator that continuously produces hard negatives, improving training stability.

The objective function adopts a pairwise ranking loss (with a margin of 1) to directly optimize ranking quality, using a squared hinge surrogate for unweighted graphs and extensions for weighted graphs.

Experimental results on OGB link‑property‑prediction benchmarks show that the pairwise ranking loss consistently outperforms classification loss, and the proposed models achieve top‑rank positions on several datasets.

Potential risk‑control applications are explored:

Self‑supervised learning via link prediction to pre‑train encoders for downstream fraud detection.

End‑to‑end relation discrimination, distinguishing trustworthy (white) and suspicious (black) relationships.

The article concludes with acknowledgments and references to the associated papers and open‑source code (https://github.com/zhitao-wang/PLNLP).