Showing 35 papers for 2026-03-10
We propose a collective learning-based graph neural network to impute missing pavement condition data by exploiting spatial-temporal correlations across the road network. The method addresses systematic missingness, reduces information loss, and improves maintenance planning decisions.
We propose a meta-learning framework for traffic assignment that generalizes across regime changes and network disruptions, leveraging graph convolutional networks to model network structure. By enabling rapid adaptation to changing conditions, the approach maintains predictive accuracy when future traffic diverges from historical patterns.
We introduce HGT-Scheduler, a deep reinforcement learning approach for the Job Shop Scheduling Problem that uses heterogeneous graph transformers to capture distinct relation types such as operation precedence and machine contention. This helps preserve important relational information that is lost when modeling the graph as homogeneous, leading to improved scheduling performance.
This study investigates how graph sparsification affects GNN pipelines at scale. It shows that selectively pruning neighbors can reduce computation with limited accuracy loss, and discusses trade-offs between performance and data movement.
We develop a dual-graph spatiotemporal GNN surrogate to predict nonlinear responses of reinforced concrete beams under four-point bending. Training data are generated from parametric Abaqus simulations by shifting loading blocks, and the model autoregressively rolls out time histories. The approach offers a computationally efficient alternative to expensive FE analyses.
We present FedShift, a two-stage distributed adversarial attack on Federated Graph Learning. In Stage 1, a hidden 'shifter' is injected into part of the training data before FedGL begins; Stage 2, the attacker seeks to degrade performance while remaining stealthy.
We propose SAGAD, a scalable and adaptive framework for graph anomaly detection that mitigates homophily disparity. SAGAD precomputes multi-hop embeddings and applies reparameterized Chebyshev filters to enable scalable detection. Additionally, it demonstrates improved accuracy and efficiency on benchmark tasks.
SCL-GNN tackles generalization in graphs by learning to ignore spurious correlations in training data. The framework identifies and decouples spurious signals during training to produce robust representations that transfer across tasks.
This study assesses how graph construction choices for tabular NetFlow data influence GNN-based IoT botnet detection performance. It compares different construction strategies and highlights their impact on long-range dependencies captured by attention-based and graph models.
We explore using graph neural networks to estimate muon momentum for CMS trigger systems. The work builds graphs from detector data and compares two graph-construction strategies, showing potential improvements in trigger efficiency.
We propose a topology-aware reinforcement learning over graphs for outage management in power distribution networks. By embedding higher-order topological features, the RL agent learns reconfiguration and load-shedding policies that maximize energy supply while preserving stability.
Rel-MOSS addresses class imbalance in relational deep learning on relational databases by modeling data as heterogeneous graphs and proposing methods to improve minority class representation.
We characterize and upgrade a quantum graph neural network for charged-particle tracking, considering LHC-scale challenges. The study analyzes QGNN performance and proposes architecture enhancements to improve tracking accuracy.
GraphProp advocates training graph foundation models by leveraging graph properties, not just node features, to improve cross-domain generalization. It argues that graph structures carry more domain-consistent information and proposes methods to exploit this.
AEGIS introduces authentic edge growth in sparsity for link prediction in edge-sparse bipartite knowledge graphs. The framework resamples existing training edges (uniform or inverse-degree biased) to avoid fabricating endpoints while preserving the original node set.
The paper analyzes how feature interactions are modeled in graph-based tabular deep learning. It argues that many GTDL methods focus on predictive accuracy while under- or mis- modeling the underlying graph structure, and uses synthetic data with known interactions to illustrate this.
We propose robust GNN verification via lightweight satisfiability testing. The approach checks for adversarial perturbations that could flip predictions and provides scalable verification for graph models.
FlowSymm presents a physics-aware, symmetry-preserving graph attention model for network flow completion. It uses a group-action on divergence-free flows, a graph-attention encoder to learn symmetry-informed weights, and a Tikhonov refinement solved via implicit bilevel optimization.
Diffusion-guided pretraining for brain graph foundation models is proposed to better preserve semantically meaningful connectome patterns. It argues against naive augmentations and introduces diffusion-based pretraining to capture global structure and improve downstream brain-graph tasks.
Benchmarking GNN models on molecular regression tasks using CKA-based representation analysis reveals how different architectures encodes representations, guiding model selection and understanding.
EpisTwin is a neuro-symbolic framework that grounds generative reasoning in a verifiable Personal Knowledge Graph to tackle fragmented personal data. It leverages multimodal language models to fuse heterogeneous, cross‑application information, enabling holistic, user‑centric sensemaking beyond traditional retrieval augmented generation.
NGDBench provides a unified benchmark for evaluating neural graph database capabilities across five domains, including finance, medicine, and AI tooling. Unlike earlier benchmarks limited to basic logical operations, NGDBench supports the full Cypher query language, enabling complex pattern matching and variable-length path queries.
SCOUT proposes a relational semantic reasoning framework that uses 3D scene graphs to support open‑world interactive object search. It addresses inefficiencies of vision–language embeddings and costly LLMs by reasoning over object relations to guide exploration, aided by a learned utility function for real‑time, context‑aware search.
Integrates a node transformer architecture with BERT-based sentiment analysis for stock price forecasting. The model represents the stock market as a graph and fuses structural price dynamics with textual sentiment signals to improve predictive performance in noisy, non-stationary markets.
MASFactory is a graph‑centric framework to orchestrate LLM‑based multi‑agent systems. It treats workflows as directed computation graphs where nodes are agents or sub‑workflows and edges encode dependencies and message passing. The framework addresses manual wiring, limited reuse, and heterogeneous context integration with reusable graph orchestration and Vibe Graphing capabilities.
Proposes a graph‑based causal framework for oppositional narrative analysis, representing narratives as entity‑interaction graphs and modeling causal relations among entities to detect and classify oppositional narratives.
Investigates ensemble strategies for probabilistic sea surface temperature forecasting using graph neural networks, focusing on how input perturbation design affects forecast skill and uncertainty representation. The study applies a homogeneous bagging‑style ensemble to the Canary Islands region to improve robustness and uncertainty quantification.
Knowledge Graphs are Implicit Reward Models: Path‑Derived Signals Enable Compositional Reasoning
Proposes a post‑training pipeline where knowledge graphs act as implicit rewards to guide compositional reasoning. The method blends supervised fine‑tuning with reinforcement learning, using path‑derived signals from KGs to train models that compose facts to solve unseen tasks.
Provides a geometric explanation for why GNN solvers struggle on harder SAT instances by analyzing graph Ricci curvature (RC). The authors show that bipartite graphs derived from random k-SAT formulas exhibit negative curvature, correlating with optimization difficulty, offering a fundamental limitation insight.
Introduces a principled hybrid model that couples a dynamic Graph Attention Network (GATv2) with a latent Gaussian spatial process from MBG. The approach decouples relational (graph) structure from continuous spatial dependence to improve spatio‑temporal risk mapping and uncertainty quantification.
Introduces DS‑DGA‑GCN, a diversity‑ and similarity‑aware dynamic graph attention‑enhanced graph convolutional network for detecting fake reviewer groups in evolving networks. It addresses cold‑start scenarios with sparse data by leveraging adaptive graph learning and temporal dynamics to identify organized, deceptive cohorts.
Proposes scenario‑aware hypergraph learning to improve next POI recommendations in LBSNs. It captures mobility variation across scenarios (e.g., tourists vs locals) and builds a hypergraph that models higher‑order relations, addressing limitations of sequential or flat‑graph methods.
Revisits retrieval for GraphRAG frameworks, proposing path‑aware subgraph selection to obtain minimal yet sufficient reasoning contexts. The approach uses graph foundation models to avoid heuristic or domain‑specific biases, improving performance in data‑scarce, cold‑start domains.
Proposes PCFEx, novel techniques to extract point-, edge-, and graph-level features from 3D point clouds treated as graphs for tasks such as human pose estimation and action recognition. It also introduces a GNN architecture optimized for efficiently processing these multi‑level features.