Reference

Structural link candidates

Behavior

For neighborhoods Γ(u) and Γ(v):

Score	Calculation	Tendency
Common neighbors	`	Γ(u) ∩ Γ(v)
Jaccard	Intersection divided by union	Normalizes neighborhood size
Adamic–Adar	Sum of 1 / log(degree(w))	Gives rare shared neighbors more weight
SimRank	Recursive similarity of incoming neighborhoods	Finds similarity through the wider incoming structure

How it works

The functions score candidate pairs supplied by the caller. Edge creation remains a caller decision. The caller also interprets each score. Direction and edge types are expressed by the neighbors callback. It can apply temporal filters and authorization before returning adjacent nodes.

Rank application-approved candidate pairs

from mari_components.graph import score_link_candidates

scores = score_link_candidates(
    candidate_pairs=(("alice", "project-x"), ("alice", "project-y")),
    neighbors=authorized_neighbors,
    method="adamic_adar",
)

for candidate in scores:
    if candidate.score >= review_threshold:
        review_queue.add(candidate)

simrank_scores is available separately because it scores all pairs in a bounded node set and iterates over incoming-neighbor similarity. It is substantially more expensive than local scores.

Keep candidate endpoints and all returned neighbors inside the authorized time-sliced graph. Remove held-out links before calculating neighborhoods for an evaluation. Jaccard and common-neighbor scores depend directly on that topology, so including the target relation can leak the answer into the score. Cache scores against graph membership and revision inputs through the dependency planner.

Measures

Split	Measure
Held-out observed links	Hits@k, MRR, ROC-AUC, average precision
Time-based split	Future-link precision with a cutoff at evaluation time
Candidate generator	Recall before structural scoring
Degree slices	Performance for sparse and hub nodes separately

Papers and implementations

Link prediction problemAdamic–AdarSimRankNetworkX link prediction

Scores provide topological evidence. The caller assigns semantic meaning.