Showing 30 papers for 2026-07-23
We introduce STN-TGAT, a model that merges a temporal Transformer with a graph attention network to capture long-horizon temporal patterns and dynamic cross-stock dependencies for top-K portfolio construction. The method uses an NMI-based prior graph and a learnable soft-threshold sparsification to improve robustness by reducing spurious connections.
This work treats arsenic concentration prediction as a regression task and builds a spatially integrated dataset to aggregate data over tens of thousands of locations. A Graph Neural Network is used to model spatial dependencies and regional heterogeneity to predict groundwater arsenic levels, aiming for generalizable spatial predictions across regions.
We propose Adaptive Multi-Expert Graph Transformer for EEG diagnostics. It treats each EEG as a sequence of dynamic functional connectivity graphs, computes wPLI-based connectivity and uses hierarchical graph encoding from electrode to regional/global levels. A multi-expert transformer aggregates outputs for interpretable predictions.
The paper conducts a controlled study to understand when knowledge graphs help RL, varying task, KG injection mechanism (state features, action masking, or potential-based reward shaping), and KG quality. It identifies conditions under which KG structure is beneficial, neutral, or detrimental to learning.
FedLSG introduces LLM-enhanced semantic calibration for federated graph backdoor defense. It grounds local graph structures and client updates into text representations, enabling semantic understanding to detect and mitigate backdoor threats in FedGNNs.
We propose label-free finite-volume-residual training for attention GNNs to simulate coupled thermo-fluid fields. Training minimizes FVM residuals computed on the mesh, requiring no labeled data, and the resulting surrogates are evaluated against CFD benchmarks.
AuditVotes proposes a provable robust defense for GNNs by combining graph rewiring augmentation and conditional smoothing within a randomized smoothing framework. It aims to achieve high clean accuracy while providing strong certified robustness.
The paper introduces a cellular sheaf-theoretic framework for complex network modeling and attention in GNNs/TDL, analyzing how signals diffuse and how training shapes feature distribution on graphs and complexes. It provides theoretical insights for topological deep learning.
AdvSynGNN offers structure-adaptive GNNs that withstand structural noise and heterophily via adversarial synthesis and self-corrective propagation. It uses multi-resolution structural synthesis and contrastive objectives to initialize geometry-aware representations and adapt to heterophilous graphs.
Boundary Embedding Shaping uses adaptive contrastive learning to address graph structural entanglement near decision boundaries. It focuses on nodes near class boundaries where structural noise most harms predictions, shaping embeddings to improve robustness.
ChemHyperMag builds a functional-group hypergraph to capture asymmetric interactions and nonreversible dynamics for multitask ADMET with missing labels. It leverages physics-informed relations among chemical motifs to improve predictive performance.
Geometry-Guided Generative Representation argues that dense functional brain graphs lie on a low-dimensional latent geometry. The paper proposes generative representations that respect this geometry, combining topological and spectral features to model brain connectomes.
Edge-Local and Qubit-Efficient Quantum Graph Learning introduces a fully quantum graph convolutional architecture designed for unsupervised learning on NISQ hardware, combining a variational quantum feature extractor with edge-local processing to reduce circuit depth and qubit requirements.
CityGuard presents a graph-aware private descriptor approach for bias-resilient identity search across urban cameras. It uses a dispersion-adaptive metric learner and topology-aware transformer to preserve privacy while improving re-identification under domain shifts.
ProbSPARQL extends SPARQL to query knowledge graphs with multi-dimensional, uncertain numeric data. It introduces probabilistic uncertainty representation for numeric values to support downstream triage, reliability modeling, and planning.
TRACER combines trajectory-aware clinical data with knowledge graphs and retrieval-augmented generation to predict clinical risk. It constructs a trajectory-aware, clinically grounded framework that leverages external knowledge to improve risk prediction.
Neuro-Symbolic Meta-Policies for Temporal Knowledge-Graph Memory proposes a neuro-symbolic meta-policy to decide which symbolic memory heuristic to apply in partially observable RL. It uses temporal RDF-annotated memory within RoomKG, combining KG encoding with memory retrieval.
Attacking Graph Foundation Models Through Their Shared Representation analyzes the alignment layer of graph foundation models as an attack surface accessible at inference time, demonstrating perturbations in representation space can break model behavior.
OntoBook converts medical ontology into pretraining signals by performing random walks, generating prose with an LLM, and training a French encoder with masked language modeling and relation prediction objectives. It shows gains on medical text tasks.
OpenRTAG provides a comprehensive benchmark for robust text-attributed graph learning under data quality degradation across nine degradation scenarios, evaluating existing methods and providing datasets/tasks for TAG robustness.
The authors show that incorporating multiple relation types with multi-relational GNNs yields richer user representations for sequential learning tasks. They identify the limitations of homogeneous relation treatment in traditional GCNs and propose a MR-GCN based framework for sequential learner modeling. Experiments on relevant tasks demonstrate improved predictive performance and more informative user models.
We study whether a GNN can approximate Subtree Prune and Regraft (SPR) distances between phylogenetic trees in near-constant time after training. The authors build and publicly release a dataset of 864 trees inferred by UPGMA and Neighbor-Joining across four bacterial species, and train a GNN to estimate SPR distance. Results show that the model provides accurate approximations quickly, enabling scalable comparisons for large phylogenies.
We propose an auditable, layered fraud detection pipeline that combines a gradient-boosted classifier, graph-derived structural features, an autoencoder anomaly signal, TreeSHAP explanations, and a bounded LLM investigation agent for uncertain cases. We also identify and remove a simulator-specific balance shortcut to prevent inflated baselines and evaluate the approach on PaySim data, demonstrating improved explainability and auditability alongside competitive performance.
Standard GNNs rely on transductive feature initializations tied to a fixed network, hindering transfer to new urban environments. We propose a network-agnostic initialization layer that guides feature initialization independently of the underlying network, improving spatial transferability of GNN-based traffic models. Experiments show the approach enables faster adaptation and better performance when transferring to unseen cities.
We propose an evidence-grounded framework that jointly predicts supply chain bottlenecks and generates faithful natural-language explanations by coupling a Temporal Graph Attention Network (TGAT) with a reasoning module based on LLM. Using maritime hubs as a case study, the system provides early warnings of bottlenecks and corresponding explanations, improving interpretability and operational usefulness.
GSPRec argues that item representations in graph signal processing based collaborative filtering should incorporate proximity information beyond the user-item interaction matrix. By capturing intermediate-frequency components corresponding to item neighborhoods derived from user sequences, GSPRec improves item embeddings. Experiments show improved recommendation accuracy and diversity.
ChiGAD introduces a spectral GNN for heterogeneous information networks using Chi-square wavelets to capture multi-path semantics and preserve high-frequency signals. It addresses abnormal signal extraction across diverse meta-paths, maintains high-frequency content for better anomaly discrimination, and learns effectively from imbalanced anomaly data. Experimental results on heterogeneous GAD benchmarks show improved detection performance over baselines.
We extend CPGRec with a balance-oriented framework that explicitly trades off accuracy and diversity in personalized video game recommendations. The approach accounts for uneven significance of player-game interactions, addressing data sparsity and potential over-smoothing. Experiments show better coverage and user satisfaction while maintaining competitive accuracy.
This paper proposes modeling scuba diving groups as temporal social networks, with divers as nodes and interactions as links. The framework focuses on three interaction dimensions—physical distance, underwater communication effectiveness, and emergency distance—to capture spatial positioning and safety dynamics. This enables computational social science analyses of dive groups.
Generative recommendation reframes sequential recommendation as autoregressive generation, but quantization of item embeddings distorts topology and misleads similarity. We propose topology-aware tokenization to preserve intrinsic item adjacency during discretization, aligning token space with the original topology. Experiments show improved generation accuracy and recommendation quality.