Link prediction

The relationalstats.linkprediction module provides link prediction metrics for NetworkX graphs.

The first public implementation includes proxfun_full, an extensible Python implementation inspired by R linkprediction::proxfun.

Public API

from relationalstats.linkprediction import ProxFun, ProxFunResult, proxfun_full

Functional API:

scores = proxfun_full(
    G,
    pairs=None,
    metrics=["jaccard", "adamic_adar", "resource_allocation"],
)

Estimator-style API:

result = ProxFun(metrics=["jaccard", "adamic_adar"]).fit(G)

result.scores_
result.to_dataframe()
result.top_k(k=20, metric="jaccard")

Documentation

• Metric formulas
• proxfun_full API notes
• ProxFunResult API notes
• Manual small-graph tests
• Internal refactor notes
• Scalability notes
• Validation against R

Scope

This module focuses on practical link prediction metrics for small and medium networks, with sparse matrix support for several local metrics.

Global metrics such as Katz, random walk with restart, and average commute time may be expensive for large networks.

Methodological note

Metric definitions can vary across packages. Validation against R linkprediction::proxfun should always align:

• graph direction;
• node ordering;
• pair selection;
• zero-division behavior;
• global metric parameters;
• disconnected-pair handling.