DyGFM introduces decoupled pre-training and divergence-conditioned prompts to create the first multi-domain dynamic graph foundation model that outperforms baselines on node classification and link prediction.
hub
Temporal Graph Networks for Deep Learning on Dynamic Graphs
22 Pith papers cite this work. Polarity classification is still indexing.
22
Pith papers citing it
abstract
Graph Neural Networks (GNNs) have recently become increasingly popular due to their ability to learn complex systems of relations or interactions arising in a broad spectrum of problems ranging from biology and particle physics to social networks and recommendation systems. Despite the plethora of different models for deep learning on graphs, few approaches have been proposed thus far for dealing with graphs that present some sort of dynamic nature (e.g. evolving features or connectivity over time). In this paper, we present Temporal Graph Networks (TGNs), a generic, efficient framework for deep learning on dynamic graphs represented as sequences of timed events. Thanks to a novel combination of memory modules and graph-based operators, TGNs are able to significantly outperform previous approaches being at the same time more computationally efficient. We furthermore show that several previous models for learning on dynamic graphs can be cast as specific instances of our framework. We perform a detailed ablation study of different components of our framework and devise the best configuration that achieves state-of-the-art performance on several transductive and inductive prediction tasks for dynamic graphs.
hub tools
citation-role summary
background 4
citation-polarity summary
roles
background 4representative citing papers
ATLAS achieves 12-30x faster out-of-core full-graph GNN inference on graphs up to 4B edges by switching to broadcast-based layer-wise execution with graph reordering, minimum-pending-message eviction, and GPU-accelerated tiered memory-disk hierarchy.
Event-level Shapley and feature-level Owen-value explainers for TGNNs outperform prior methods on metrics and datasets while revealing a timestamp extraction bug in TGAT.
TravelFraudBench is a new configurable benchmark for GNN-based fraud ring detection in travel networks, simulating star, clique, and chain topologies and showing GraphSAGE outperforming MLP baselines on AUC and ring recovery.
PROVFUSION fuses three complementary views of provenance data with lightweight schemes and voting to achieve higher detection accuracy and lower false positives than node- or edge-only baselines on nine benchmarks.
ChronoSpike is a spiking GNN that integrates adaptive LIF neurons with spatial attention and temporal transformers to outperform baselines on dynamic graph benchmarks by 2% F1 while training 3-10x faster with fixed parameters and stability guarantees.
Unsupervised GNN model learns local updates for approximate MaxIS on dynamic graphs, achieving competitive ratios on 200-1000 node instances and 1.00-1.18x larger solutions than other unsupervised models when generalizing to 100x larger graphs.
Graph Retention Networks extend retention to dynamic graphs to enable parallelizable training, O(1) inference, and chunkwise long-term training while delivering competitive performance with major efficiency gains.
Diagnoses attention dispersion in CTDG Transformers under temporal shift and introduces differential attention to suppress common signals and achieve SOTA on shifted benchmarks.
DRIFT benchmark shows substantial performance degradation for continual graph learning methods under task-free continuous distribution shifts modeled via Gaussian mixtures.
Simple MLPs using temporal and behavioral features from gossip data predict Lightning Network channel closure types better than temporal graph neural networks.
FAME models scientific topic trajectories in continuous time to forecast paper impact more accurately than LLMs by aligning manuscripts with field momentum in a dynamic latent space.
PRISM iteratively transforms semantic priors into behavior-conditioned posteriors via cross-modal refinement to improve representation learning on dynamic text-attributed graphs.
OKH-RAG represents knowledge as ordered hyperedges and retrieves coherent interaction sequences via a learned transition model, outperforming permutation-invariant RAG baselines on order-sensitive QA tasks.
BiTA redesigns temporal aggregation in TGNs by jointly using bidirectional GRU for sequential dependencies and Transformer for long-range context to improve alert prediction accuracy on real network data.
LLM multi-agent systems augmented with data-driven event triggers and Hawkes processes simulate both micro-level interactions and macroscopic topologies in dynamic email networks for realistic phishing synthesis.
Introduces consistent event graph isomorphism and a temporal Weisfeiler-Leman algorithm to analyze and improve the expressive power of message passing in temporal event graphs.
Geometric deep learning provides a unified mathematical framework based on grids, groups, graphs, geodesics, and gauges to explain and extend neural network architectures by incorporating physical regularities.
A2QTGN combines adaptive quantum amplitude encoding with a temporal graph network to improve dynamic link prediction, showing strong results on five benchmark datasets.
Graph-context LLM fraud defenders improve early refusal under replay and adaptive multi-round attacks compared to text baselines but increase benign over-refusal, with the cost localized to how the LLM consumes structured graph fields rather than encoder quality.
ST-GAT applies spatial-temporal graph attention networks to reconstructed interbank graphs from FDIC Call Reports, achieving 0.939 AUPRC for bank distress prediction with explainable feature importance.
AdvSynGNN uses multi-resolution structural synthesis, contrastive objectives, an adaptive transformer, and an adversarial propagation engine with residual label correction to improve node-level predictions on challenging graph topologies.
citing papers explorer
-
Decoupled and Divergence-Conditioned Prompt for Multi-domain Dynamic Graph Foundation Models
DyGFM introduces decoupled pre-training and divergence-conditioned prompts to create the first multi-domain dynamic graph foundation model that outperforms baselines on node classification and link prediction.
-
ATLAS: Efficient Out-of-Core Inference for Billion-Scale Graph Neural Networks
ATLAS achieves 12-30x faster out-of-core full-graph GNN inference on graphs up to 4B edges by switching to broadcast-based layer-wise execution with graph reordering, minimum-pending-message eviction, and GPU-accelerated tiered memory-disk hierarchy.
-
Explaining Temporal Graph Predictions With Shapley Values
Event-level Shapley and feature-level Owen-value explainers for TGNNs outperform prior methods on metrics and datasets while revealing a timestamp extraction bug in TGAT.
-
TRAVELFRAUDBENCH: A Configurable Evaluation Framework for GNN Fraud Ring Detection in Travel Networks
TravelFraudBench is a new configurable benchmark for GNN-based fraud ring detection in travel networks, simulating star, clique, and chain topologies and showing GraphSAGE outperforming MLP baselines on AUC and ring recovery.
-
Beyond Nodes vs. Edges: A Multi-View Fusion Framework for Provenance-Based Intrusion Detection
PROVFUSION fuses three complementary views of provenance data with lightweight schemes and voting to achieve higher detection accuracy and lower false positives than node- or edge-only baselines on nine benchmarks.
-
ChronoSpike: An Adaptive Spiking Graph Neural Network for Dynamic Graphs
ChronoSpike is a spiking GNN that integrates adaptive LIF neurons with spatial attention and temporal transformers to outperform baselines on dynamic graph benchmarks by 2% F1 while training 3-10x faster with fixed parameters and stability guarantees.
-
Unsupervised Learning of Local Updates for Maximum Independent Set in Dynamic Graphs
Unsupervised GNN model learns local updates for approximate MaxIS on dynamic graphs, achieving competitive ratios on 200-1000 node instances and 1.00-1.18x larger solutions than other unsupervised models when generalizing to 100x larger graphs.
-
Graph Retention Networks for Dynamic Graphs
Graph Retention Networks extend retention to dynamic graphs to enable parallelizable training, O(1) inference, and chunkwise long-term training while delivering competitive performance with major efficiency gains.
-
Attention Dispersion in Dynamic Graph Transformers: Diagnosis and a Transferable Fix
Diagnoses attention dispersion in CTDG Transformers under temporal shift and introduces differential attention to suppress common signals and achieve SOTA on shifted benchmarks.
-
DRIFT: A Benchmark for Task-Free Continual Graph Learning with Continuous Distribution Shifts
DRIFT benchmark shows substantial performance degradation for continual graph learning methods under task-free continuous distribution shifts modeled via Gaussian mixtures.
-
Predicting Channel Closures in the Lightning Network with Machine Learning
Simple MLPs using temporal and behavioral features from gossip data predict Lightning Network channel closure types better than temporal graph neural networks.
-
FAME: Forecasting Academic Impact via Continuous-Time Manifold Evolution
FAME models scientific topic trajectories in continuous time to forecast paper impact more accurately than LLMs by aligning manuscripts with field momentum in a dynamic latent space.
-
PRISM: Iterative Cross-Modal Posterior Refinement for Dynamic Text-Attributed Graphs
PRISM iteratively transforms semantic priors into behavior-conditioned posteriors via cross-modal refinement to improve representation learning on dynamic text-attributed graphs.
-
Knowledge Is Not Static: Order-Aware Hypergraph RAG for Language Models
OKH-RAG represents knowledge as ordered hyperedges and retrieves coherent interaction sequences via a learned transition model, outperforming permutation-invariant RAG baselines on order-sensitive QA tasks.
-
BiTA: Bidirectional Gated Recurrent Unit-Transformer Aggregator in a Temporal Graph Network Framework for Alert Prediction in Computer Networks
BiTA redesigns temporal aggregation in TGNs by jointly using bidirectional GRU for sequential dependencies and Transformer for long-range context to improve alert prediction accuracy on real network data.
-
Can LLM Agents Simulate Dynamic Networks? A Case Study on Email Networks with Phishing Synthesis
LLM multi-agent systems augmented with data-driven event triggers and Hawkes processes simulate both micro-level interactions and macroscopic topologies in dynamic email networks for realistic phishing synthesis.
-
Weisfeiler and Leman Follow the Arrow of Time: Expressive Power of Message Passing in Temporal Event Graphs
Introduces consistent event graph isomorphism and a temporal Weisfeiler-Leman algorithm to analyze and improve the expressive power of message passing in temporal event graphs.
-
Geometric Deep Learning: Grids, Groups, Graphs, Geodesics, and Gauges
Geometric deep learning provides a unified mathematical framework based on grids, groups, graphs, geodesics, and gauges to explain and extend neural network architectures by incorporating physical regularities.
-
A2QTGN: Adaptive Amplitude Quantum-Integrated Temporal Graph Network for Dynamic Link Prediction
A2QTGN combines adaptive quantum amplitude encoding with a temporal graph network to improve dynamic link prediction, showing strong results on five benchmark datasets.
-
Rethinking Fraud Safety Evaluation: Multi-Round Attacks Reveal Safety-Utility Tradeoffs in Graph-Context LLM Defenders
Graph-context LLM fraud defenders improve early refusal under replay and adaptive multi-round attacks compared to text baselines but increase benign over-refusal, with the cost localized to how the LLM consumes structured graph fields rather than encoder quality.
-
Explainable Graph Neural Networks for Interbank Contagion Surveillance: A Regulatory-Aligned Framework for the U.S. Banking Sector
ST-GAT applies spatial-temporal graph attention networks to reconstructed interbank graphs from FDIC Call Reports, achieving 0.939 AUPRC for bank distress prediction with explainable feature importance.
-
AdvSynGNN: Structure-Adaptive Graph Neural Nets via Adversarial Synthesis and Self-Corrective Propagation
AdvSynGNN uses multi-resolution structural synthesis, contrastive objectives, an adaptive transformer, and an adversarial propagation engine with residual label correction to improve node-level predictions on challenging graph topologies.