Link Prediction

Classical approaches that compute connection likelihood based on topological proximity without requiring model training.

• Common Neighbors: Scores node pairs by the number of shared adjacent nodes, assuming cases cited by the same authorities are related.
• Jaccard Coefficient: Normalizes common neighbors by total neighborhood size to avoid bias toward high-degree nodes.
• Adamic-Adar: Weights shared neighbors inversely by their degree, prioritizing rare connections over ubiquitous citations.
• Preferential Attachment: Predicts links based on the product of node degrees, modeling the tendency of highly-cited cases to attract future citations.

Methods that learn low-dimensional vector representations of legal nodes by preserving graph structure, enabling similarity-based link prediction.

• Node2Vec: Uses biased random walks to capture both homophily and structural equivalence in citation patterns.
• DeepWalk: Treats truncated random walks as sentences, applying skip-gram models to learn latent representations of cases.
• Spectral Embedding: Decomposes the graph Laplacian matrix to position nodes in a space where connected cases are proximate.
• Embedding distance or dot product between two case vectors serves as the link probability score.

Deep learning architectures that generate node representations by recursively aggregating features from local citation neighborhoods.

• Graph Convolutional Networks (GCNs) apply spectral convolutions to learn node embeddings that capture multi-hop citation dependencies.
• GraphSAGE samples and aggregates features from a node's local neighborhood, enabling inductive prediction for newly published cases.
• Graph Attention Networks (GATs) assign learnable importance weights to different citing authorities during neighborhood aggregation.
• An edge decoder—typically a multi-layer perceptron or dot product—computes the final link probability from the learned node pair embeddings.

Techniques that incorporate the time dimension to predict future citations rather than merely recovering missing historical links.

• Temporal Edge Features: Encodes the age difference between cases and the recency of prior citations as input signals.
• Time-Aware Negative Sampling: Selects negative edges from the same time window as positive edges to prevent temporal leakage during training.
• Recurrent Graph Architectures: Combines GNNs with RNNs or temporal attention to model the evolution of citation networks over sequential time snapshots.
• Hawkes Process Integration: Models citation events as a self-exciting point process where each citation temporarily increases the probability of subsequent citations.

Quantitative measures used to assess link prediction performance, critical given the extreme class imbalance in sparse citation graphs.

• Area Under ROC Curve (AUC-ROC) measures the trade-off between true positive and false positive rates across thresholds.
• Mean Reciprocal Rank (MRR) evaluates how highly the correct missing link is ranked among all candidate edges.
• Hits@K measures the fraction of true links appearing in the top-K predictions, with K values of 10, 50, or 100 being standard.
• Evaluation requires temporal splitting: training on citations before a cutoff date and testing on citations formed after that date to simulate real-world forecasting.

Domain-specific input signals that enhance link prediction beyond pure topology by encoding jurisprudential knowledge.

• Court Hierarchy Distance: Encodes the vertical separation between the issuing courts of two cases within the judicial system.
• Subject Matter Overlap: Measures the semantic similarity of legal topics or West Key Number classifications between cases.
• Treatment Sentiment Features: Incorporates the polarity of existing citations as edge weights, distinguishing supportive from critical authority chains.
• Jurisdictional Concordance: A binary or scalar feature indicating whether two cases originate from the same sovereign jurisdiction, a strong predictor of citation likelihood.

A directed network structure where nodes represent legal cases or statutes and edges represent citation relationships. This forms the foundational data structure for computational precedent analysis.

• Nodes: Individual legal authorities (cases, statutes, regulations)
• Edges: Directed citations from a citing case to a cited authority
• Edge attributes: May encode treatment type, citation intent, or temporal metadata
• Scale: Modern legal graphs contain millions of nodes and tens of millions of edges

Citation graphs enable graph traversal algorithms to map the flow of legal reasoning across jurisdictions and time periods.

A graph algorithm that iteratively distributes precedential influence scores across a citation network, often using PageRank variants, to identify the most legally significant nodes.

• Mechanism: Nodes pass authority weight to their out-neighbors proportionally
• Convergence: Algorithm iterates until score distribution stabilizes
• Damping factor: Prevents rank sinks and models random judicial discovery
• Weighted variants: Incorporate treatment sentiment and court hierarchy into edge weights

Authority propagation surfaces landmark cases that may not have the highest raw citation count but sit at critical junctures in the legal reasoning network.

An NLP task that automatically categorizes how a citing case legally treats a cited authority, assigning labels such as:

• Overruled: Explicit invalidation of the prior holding
• Distinguished: Material factual or legal differences found
• Followed: Precedent applied as controlling authority
• Criticized: Negative commentary without explicit overruling
• Harmonized: Reconciled with apparently conflicting authority

Accurate treatment classification is essential for maintaining citation graph integrity and computing reliable authority scores that reflect actual judicial reception rather than mere citation frequency.

A deep learning architecture designed to operate directly on graph-structured data, used in legal AI to learn node embeddings that capture both a case's intrinsic features and its citation neighborhood structure.

• Message passing: Nodes aggregate feature information from their neighbors iteratively
• Node embeddings: Dense vector representations encoding both content and graph position
• Applications: Link prediction, node classification, and authority score refinement
• Architectures: Graph Convolutional Networks (GCNs), Graph Attention Networks (GATs), and GraphSAGE

GNNs enable link prediction models to learn complex, non-linear patterns in citation behavior that simple heuristic metrics cannot capture.

The study of citation patterns over time to model how legal authority evolves, ages, or gains influence, incorporating timestamps into graph models to detect trends like precedent aging.

• Citation velocity: Rate at which a decision accumulates citations post-publication
• Precedent aging: Decline in citation frequency as doctrines mature or are superseded
• Temporal link prediction: Forecasting future citations based on historical patterns
• Epoch encoding: Time-aware node embeddings that capture a case's historical context

Temporal analysis distinguishes between a case that is historically significant but no longer actively cited and one experiencing a resurgence in relevance due to new doctrinal developments.

A graph metric measuring how often a node lies on the shortest path between other nodes, used in citation networks to identify cases that serve as critical bridges connecting distinct doctrinal clusters.

• Computation: Fraction of all shortest paths in the graph that pass through the node
• Interpretation: High betweenness cases connect otherwise disconnected areas of law
• Bridge detection: Identifies cases that synthesize multiple doctrinal threads
• Vulnerability analysis: Removal of high-betweenness nodes fragments the citation graph

Cases with high betweenness centrality often represent seminal decisions that unified previously fragmented areas of jurisprudence, making them critical targets for link prediction models.