Skip to content

Heterogeneous Graph Neural Networks

Heterogeneous graph neural networks (HGNNs) extend standard GNNs to operate on graphs containing multiple types of nodes and edges. While homogeneous GNNs assume all nodes and edges are of the same type (e.g., all users connected by "follows" relationships), heterogeneous graphs can model richer relational structures: nodes may represent different entity types (users, documents, organizations), and edges may capture different relationship types (coauthorship, citation, funding).

Motivation

Real-world networks are often heterogeneous. In social media, a news feed contains users, posts, and accounts with multiple types of relationships (friendship, retweet, quote, mention). In academia, knowledge graphs have papers, authors, venues, and topics connected by authorship, citation, publication, and topic edges. Applying homogeneous GNNs to such networks requires flattening this structure, losing important semantic information about relationship types. HGNNs preserve and exploit type information.

Core concepts

Node and edge types: Each node and edge has an explicit type label. A heterogeneous graph is a tuple G = (V, E, T_v, T_e) where T_v maps nodes to types and T_e maps edges to types.

Type-aware aggregation: Instead of aggregating uniformly over all neighbors, HGNNs aggregate separately for each edge type or node type, then combine results. This allows the model to learn different interaction strengths for different relationship types.

Meta-paths: A common abstraction for reasoning about paths in heterogeneous graphs. A meta-path defines a sequence of node/edge types (e.g., author → paper → author describes a co-authorship relationship via shared papers). HGNNs often learn separate embeddings for paths of different types.

Applications to fake news detection

Heterogeneous graph structures naturally arise in misinformation contexts:

  • Speaker-news graphs: Nodes are speakers and news items; edges represent speaker credibility, speaker-statement relationships, and news similarity. Different edge types encode different signals: speaker reputation, party affiliation, historical accuracy.

  • User-content-propagation graphs: Nodes include users, content items, and propagation sequences; edge types distinguish initial posting, sharing, liking, and commenting behaviors. Each type may predict different aspects of credibility.

Key papers using heterogeneous GNNs for fake news detection:

A Decision-Based Heterogeneous Graph Attention Network for Multi-Class Fake uses a heterogeneous graph of the LIAR dataset where news items are connected by nine edge types derived from speaker profiles (speaker, context, subject, party-affiliation, job-title, state) and content similarity (KNN-5, KNN-6, KNN-7). A decision network learns which edge types are optimal for each node at each layer, enabling dynamic and task-specific neighborhood selection.