Skip to content

Graph Neural Networks

Graph neural networks (GNNs) are neural network architectures that learn representations of graph-structured data by aggregating information from node neighborhoods. Unlike traditional neural networks designed for fixed-size inputs (images, sequences), GNNs operate on variable-size graphs where nodes can have different numbers of neighbors and connections.

Core concepts

Message passing: GNNs typically work by iteratively updating node representations by aggregating information from neighboring nodes (message passing). Different GNN variants (Graph Convolutional Networks, GraphSAGE, Graph Attention Networks) differ in their aggregation functions and update rules.

Graph structure preservation: GNNs aim to learn representations that preserve graph properties—nodes with similar neighborhood structures or roles should have similar representations. This is useful for both node-level tasks (classification, link prediction) and graph-level tasks (predicting properties of entire graphs).

Scalability challenge: A key research question is how to scale GNNs to very large graphs (millions/billions of nodes) where computing global representations requires careful subsampling and aggregation strategies.

Applications in cascade prediction and information diffusion

GNNs are particularly useful for cascade prediction because cascades are naturally represented as graphs (nodes = users, edges = information propagation). Rather than hand-engineering features of cascade graphs (clustering coefficients, centrality measures), GNNs can automatically learn which structural properties are most predictive.

DeepCas: an End-to-end Predictor of Information Cascades represents cascades as random walk paths through the cascade graph, processed through bidirectional GRU networks with attention—an early approach to end-to-end learning of cascade representations.

Key papers

  • A Decision-Based Heterogeneous Graph Attention Network for Multi-Class Fake — DHGAT: Decision-based Heterogeneous Graph Attention Network for multi-class fake news detection. Enables per-node per-layer dynamic selection of optimal neighborhood types using a decision network and Gumbel-Softmax sampling. Represents news as heterogeneous graphs with multiple edge types from speaker profiles; achieves ~4% accuracy improvement on LIAR (12,836 items, six classes) with robustness in label-scarce regimes (10%–30% labeled).
  • Neighborhood-Order Learning Graph Attention Network for Fake News Detection — NOL-GAT: Addresses the fixed L-hop neighborhood limitation in standard GNNs by allowing each node to adaptively learn its optimal neighborhood order at each layer via Gumbel-Softmax. Applied to fake news detection with two-component architecture (Hop Network for order selection, Embedding Network for representation updates); achieves strong performance on five datasets, particularly in low-label settings.
  • The Evolution of Distributed Systems for Graph Neural Networks and their Origin — Survey of distributed systems for scalable GNN training, covering partitioning, sampling, communication, synchronization, and programming abstractions; bridges graph processing systems (Pregel, PowerGraph) and DNN training systems.
  • A Comprehensive Survey on Trustworthy Graph Neural Networks: Privacy, Robustness, — Comprehensive survey of trustworthy graph neural networks across privacy, robustness, fairness, and explainability dimensions; catalogs attacks and defenses including differential privacy, federated learning, adversarial robustness, and model ownership verification.
  • Advancing Graph Representation Learning with Large Language Models: A Comprehensive — Comprehensive survey of integrating LLMs with graph representation learning and GNNs; breaks down techniques into knowledge extractors (attribute, structure, label) and organizers (GNN-centric, LLM-centric, GNN+LLM); analyzes integration strategies and training approaches for enhanced graph learning
  • DELL: Generating Reactions and Explanations for LLM-Based Misinformation Detection (2024) — Uses GNNs to encode LLM-generated user-news interaction networks and apply explainable proxy tasks for misinformation detection; demonstrates integration of GNNs with LLM-generated synthetic data and multi-task learning for improved detection on fake news, framing, and propaganda tasks.
  • Hu, Hu & Zhang (2025) — Synergizing LLMs with Global Label Propagation for Multimodal Fake News Detection — Combines LLM-generated pseudo labels with GCN-based label propagation over cross-modal graphs; introduces mask-based global label propagation to prevent label leakage while enabling effective information propagation across multimodal samples.
  • Dou et al. (2021) — User Preference-aware Fake News Detection (UPFD): Applies GNNs (GraphSAGE, GCN) to jointly encode user endogenous preferences and exogenous propagation context. Uses text embeddings (BERT/word2vec) as node features in news propagation graphs; readout aggregation produces user engagement embedding; concatenated with news textual embedding for classification. Demonstrates GNN-based integration of heterogeneous signals (user behavior + graph structure) outperforms single-signal baselines.
  • Lu & Li (2020) — GCAN: Graph-aware Co-Attention Networks: Uses GCN to model pairwise user interactions in retweet sequences; couples graph-aware user representations with dual co-attention mechanism; achieves 87.7% accuracy on Twitter15 and 90.8% on Twitter16 for fake news detection; provides explainability via attention weight visualization.
  • Rumor Detection on Social Media with Bi-Directional Graph Convolutional Networks — First GCN application to rumor detection; combines bidirectional graph convolutions (top-down propagation + bottom-up dispersion) with source-post feature enhancement; achieves state-of-the-art rumor detection on three benchmarks (96.1% Weibo, 88.6% Twitter15, 88.0% Twitter16).
  • Bridging Thoughts and Words: Graph-Based Intent-Semantic Joint Learning for — InSIDE: Heterogeneous graph learning combining semantic and intent signals for fake news detection. Models news as two interconnected graphs: G^sem (sentences, entities) with local narrative flow and global entity-mediated context; G^int (coarse-fine intent hierarchy) capturing purposes/motivations. Dynamic pathway-based alignment via pseudo nodes bridges the two signals. Validates on diverse datasets and intent frameworks.
  • SINcon: Mitigate LLM-Generated Malicious Message Injection Attack for Rumor — Defends GNN-based MPT rumor detection against LLM-generated message injection attacks via contrastive learning; identifies and mitigates vulnerability from imbalanced node influence scores; achieves 26% robustness improvement on Twitter/Weibo while preserving clean accuracy.
  • DeepCas: an End-to-end Predictor of Information Cascades — Uses random walk paths and GRU networks with attention to predict cascade size; demonstrates automated learning of cascade graph representations
  • Topological Recurrent Neural Network for Diffusion Prediction — First LSTM designed for dynamic DAGs (diffusion topologies); extends neural approaches to non-tree graph structures with changing topology over time
  • Jin et al. (2021) — Towards Fine-Grained Reasoning for Fake News Detection: Uses a Prior-Aware Bi-Channel Kernel Graph Attention Network (KGAT) to perform fine-grained reasoning over claim-evidence graphs; integrates both textual and social channels via kernel matching to achieve 91.7% F1 on PolitiFact; demonstrates how GNNs with kernel-based attention mechanisms can enable interpretable fake news detection.
  • Perceived and Intended Sarcasm Detection with Graph Attention Networks — Graph Attention Networks combining user history embeddings and heterogeneous social graphs for sarcasm detection on Twitter; demonstrates that GATs over user-tweet interaction networks can distinguish sarcastic intent from perception
  • Network Information Enhances Unreliable News Domain Detection — Domain co-sharing networks on Telegram; GraphSAGE outperforms content-unaware baselines on source reliability classification; demonstrates network topology carries complementary signals to content features for credibility assessment.