Heterogeneous Graph

Unlike homogeneous graphs that model a single entity type, a heterogeneous graph defines a schema with multiple, distinct node categories. In a legal network, this includes nodes for Case, Statute, Court, Judge, Party, and Legal Doctrine. Each node type carries a unique set of attributes—a Judge node has a name and appointment date, while a Statute node has a jurisdiction and effective date. This explicit typing enables the system to apply type-specific validation rules and prevents nonsensical connections, such as a direct edge between two judges without a shared case.

Relationships in a heterogeneous graph are first-class citizens with defined types and directionality. Legal networks require edges like CITES, AFFIRMS, OVERRULES, ENACTS, and INTERPRETS. An edge from a Case node to a Statute node typed INTERPRETS carries a fundamentally different semantic meaning than an edge typed CHALLENGES. This rich edge typing allows traversal algorithms to follow specific legal reasoning paths, such as finding all cases that have positively cited a statute without being distracted by unrelated procedural connections.

A meta-path is a predefined sequence of node and edge types that captures a specific semantic relationship. For example, the meta-path Case -[CITES]-> Case -[CITES]-> Statute identifies a chain of precedent that leads to a statutory interpretation. Another critical meta-path is Judge -[AUTHORED]-> Case -[OVERRULED_BY]-> Case -[AUTHORED]-> Judge, which reveals judicial disagreement patterns. These structured paths are essential for authority propagation algorithms and for generating features for downstream machine learning models like Graph Neural Networks.

Standard GNNs assume all nodes and edges belong to a single type, which collapses critical legal distinctions. Heterogeneous Graph Neural Networks are specifically designed to learn distinct weight matrices for each node and edge type. An HGNN processes a legal graph by:

• Applying a unique transformation to Case node features vs. Statute node features
• Aggregating messages from CITES edges differently than from OVERRULES edges
• Learning type-specific attention mechanisms that weight the importance of different neighbor types This architecture produces node embeddings that respect the graph's inherent heterogeneity, dramatically improving performance on tasks like case outcome prediction and citation recommendation.

Heterogeneous legal graphs are typically implemented using one of two paradigms:

• RDF Triplestores: Represent all data as subject-predicate-object triples, ideal for legal knowledge graphs requiring formal ontologies and SPARQL querying across linked open data
• Labeled Property Graphs: Store nodes and edges with key-value property maps, offering faster traversal for citation network analysis and native support for algorithms like PageRank Each model supports schema enforcement, ensuring that a CITES edge can only connect two Case nodes, maintaining the graph's logical integrity for downstream legal reasoning tasks.

A defining capability of heterogeneous legal graphs is the ability to dynamically filter subgraphs based on node attributes. A query can isolate all nodes and edges within a specific jurisdiction (e.g., 'Second Circuit') and a specific temporal window (e.g., '2010-2023'). This creates a contextually relevant subgraph for stare decisis modeling, ensuring that authority scores and precedent recommendations are calculated only from legally binding sources. The heterogeneous schema allows the filter to apply simultaneously across Court, Case, and Statute node types, producing a coherent, jurisdictionally-scoped view of the law.

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

• Directionality: An edge points from the citing case to the cited authority.
• Structure: Forms the substrate upon which all authority propagation algorithms operate.
• Example: Roe v. Wade is a node; Dobbs v. Jackson Women's Health Organization has a directed edge pointing to it.

An iterative graph algorithm that distributes precedential influence scores across a citation network. Often implemented using PageRank variants, it identifies the most legally significant nodes.

• Mechanism: A node's authority is a function of the quantity and quality of its inbound citations.
• Recursive Principle: Citations from highly authoritative cases confer more weight than citations from obscure ones.
• Application: Used to rank search results by legal importance rather than just semantic relevance.

An NLP task that automatically categorizes how a citing case legally treats a cited authority. Each citation instance is assigned a label.

• Positive Treatment: 'Followed,' 'Applied,' 'Affirmed.'
• Negative Treatment: 'Overruled,' 'Reversed,' 'Criticized,' 'Questioned.'
• Neutral Treatment: 'Distinguished,' 'Explained,' 'Cited.'
• Critical Signal: Negative treatment directly impacts the precedential weight of the target node.

A deep learning architecture designed to operate directly on graph-structured data. In legal AI, GNNs 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.
• Output: A dense vector representation for every case that encodes its jurisprudential context.
• Use Case: Superior to text-only models for tasks like case outcome prediction because it incorporates the network of authority.

A machine learning task that predicts the likelihood of a missing or future edge between two nodes in a graph. Applied to citation networks to forecast which precedents a court is likely to cite.

• Input: The current state of the citation graph and node features.
• Output: A probability score for potential new citation edges.
• Application: Powers citation recommendation systems for legal drafters, suggesting relevant authority they may have overlooked.

An unsupervised graph clustering technique that partitions a citation network into groups of densely interconnected cases. This often reveals distinct legal topics, circuits, or doctrinal silos.

• Doctrinal Clusters: Groups of cases coalescing around a specific legal principle.
• Jurisdictional Clusters: Natural groupings by circuit court or state.
• Algorithm Examples: Louvain method, Label Propagation, Girvan-Newman.
• Value: Enables the mapping of the organic structure of legal thought without manual taxonomy creation.