Citation Graphs as Legal Signal: Structural Features for SCOTUS Link Prediction

We present a two-layer graph construction pipeline for U.S. Supreme Court (SCOTUS) opinions that combines deterministic citation extraction with graph-structural feature engineering to predict missing legal citations and identify influential case law. Working from 4,915 SCOTUS opinions and 7,159 directed citation edges, we construct a citation graph and compare seven feature sets — ranging from text embeddings alone to combinations of common neighbours, triangle counts, Louvain community membership, entity overlap, and semantic similarity — across Logistic Regression and Random Forest classifiers. The combined graph-and-text model achieves AUC 0.851, a 10% relative improvement over a Legal-BERT text-only baseline (AUC 0.773), while PageRank and betweenness centrality recover the legal canon and doctrinal bridge cases without supervision. Our results suggest that structural citation features carry legal-domain information that text embeddings miss, with implications for precedent retrieval, legal research automation, and access to justice.