Showing 15 papers for 2026-06-29
hia-gat is a heterogeneous interaction-aware graph attention network for frame-level traffic conflict risk prediction on freeways. It constructs a per-frame graph with vehicles as nodes and two edge types: same-lane (longitudinal) and adjacent-lane (lateral), augmented with physics-informed edge features aligned to rear-end and lane-change conflict mechanisms. The problem is formulated as frame-level binary risk classification, and results on a structured benchmark show improvements over baselines.
TeRoR addresses temporal knowledge graph embedding. While TeRo is simple and efficient, it struggles to model the mapping properties of relations and temporal information. TeRoR proposes decoupled temporal rotation with a relational circular region to better capture one-to-many, many-to-one, and many-to-many mappings and richer temporal dynamics.
GNBAN introduces Graph Neural Basis Attention Networks for long-horizon forecasting over large entity sets. It scales to tens of thousands of correlated time series by sharing basis components and applying graph-based attention, enabling accurate and interpretable forecasts for large catalogs.
Beyond Sparse Supervision proposes diffusion-guided learning for few-shot graph fraud detection. It tackles sparse and imbalanced supervision and representation dilution by using diffusion processes to propagate informative signals while preserving fraud-relevant irregularities. The approach improves detection performance under scarce labels.
Directed Graph Topology Inference develops methods to infer a directed network from nodal measurements produced by linear diffusion. Observations are modeled as outputs of a graph convolutional filter with unknown coefficients, excited by arbitrarily correlated signals, and the method handles directed graphs and non-white excitations.
Benchmarking Multi-Modal Graph-based Social Media Popularity Prediction benchmarks and analyzes how combining multimodal content and temporal social interactions affects performance. The work surveys datasets, modalities, and observation windows and provides unified baselines for evaluation.
The paper systematically investigates how relational graph modularity and depth affect learning performance of neural networks. It studies mesoscale community structures and depth-related tradeoffs to provide practical guidelines for designing graph neural networks.
Estimating condition number with Graph Neural Networks proposes a fast method for estimating the condition number of sparse matrices using GNNs. It introduces graph features with O(nnz + n) complexity and presents two estimation schemes, including a decomposition approach that targets the heavier computational component.
Graph Neural Networks for Predicting Solvability of Finite Groups develops a GNN framework to classify finite groups by solvability using graph representations, including Cayley graphs. The model is trained to distinguish solvable and non-solvable groups using only graph structure and is tested on unseen groups to assess generalization.
Pulmonary Embolism Risk Stratification from CTPA and Medical Records: Vascular Graphs Are Not All You Need studies risk stratification for pulmonary embolism using medical records and biomarkers from CTPA, addressing missing blood tests. It benchmarks strategies that combine medical records with CTPA-derived biomarkers and demonstrates that vascular graphs alone are not sufficient for optimal performance.
RelBall: Relation Ball with Quaternion Rotation for Knowledge Graph Completion introduces RelBall, a method that uses a relation ball and quaternion rotations to model diverse relational patterns, including non-commutative composition, addressing non-commutativity gaps in prior models like RotatE and Rotate3D.
Ontology-Guided Evidence Path Inference for Multi-hop Knowledge Graph Question Answering proposes OPI, a framework that uses a relation-centric ontology graph to guide evidence-path search in multi-hop KGQA, mitigating the path search space explosion and ensuring semantic alignment with complex questions.
GRAFT: Biological Graph and Hypergraph Benchmarks for Linked Gene Expression and Phenotypic Trait Prediction in Arabidopsis thaliana presents benchmarks that combine graph and hypergraph representations to study linked gene expression and trait prediction in Arabidopsis, enabling reasoning over heterogeneous biological data.
KG2Cypher: Data-Centric Pipeline for Building Enterprise Text-to-Cypher Systems presents a data-centered workflow to build enterprise text-to-Cypher systems. It first constructs executable Cypher queries from observed graph facts, then uses LLMs to generate corresponding natural-language questions, and validates the resulting Text-Cypher pairs with an LLM judge and humans.
Hybrid Fact-Checking that Integrates Knowledge Graphs, Large Language Models, and Search-Based Retrieval Agents Improves Interpretable Claim Verification outlines a hybrid approach that combines knowledge graphs, LLMs, and search-based retrieval to provide interpretable and grounded claim verification, detailing a three-stage workflow for retrieval, reasoning, and justification.