Showing 39 papers for 2026-06-30
We present an industrial case study of a self-supervised learning method that uses graph neural networks to detect anomalies in cloud log data. The approach targets reducing false positives compared with static heuristic methods by exploiting relational patterns in logs.
We develop a surrogate model of the static mechanical response of short-fiber thermoplastics using Graph Neural Networks. The response is governed by mesoscale features like fiber orientation, clustering, and porosity; the GNN captures these variations to predict stiffness, damage initiation, and nonlinear deformation.
We propose Blackknife, a hard-label, query-limited black-box attack on heterogeneous Graph Neural Networks. Unlike gradient- or score-based attacks, Blackknife operates with only predicted labels and limited queries, without access to the full graph structure.
We introduce GLIP, a framework for joint pretraining of Graph Neural Networks and Large Language Models to tackle graph-level tasks. The approach fuses structural graph representations with textual information to improve graph-level predictions.
We present T3R, a method for deeper test-time adaptation of Graph Neural Networks via gradient rotation using multiple Rotograd matrices. It enables more substantial online updates with unlabeled test data, improving robustness to distribution shifts.
We develop Query-Aware Spreading Activation for multi-hop retrieval over knowledge graphs. Unlike seed-only traversal, our flow-diffusion solver with edge re-weighting yields query-aware traversal, but it requires loading the full graph into memory and iterative solving.
We introduce a stochastic graph heat modelling approach for connectivity estimation in neurophysiological data. The method uses noise-driven graph heat diffusion to estimate dynamic, directed brain connectivity, addressing limitations of non-model-based approaches.
We provide an information-geometric justification for the composite coherence measure C = sqrt(A · T) used in event-based narrative extraction, where A is angular similarity of embeddings and T is 1 minus the Jensen-Shannon distance of soft topic memberships. We offer axiomatic characterization showing how this combination arises from information geometry.
We demonstrate an autonomous LLM research loop for optimizing expert-designed Crystal Graph Networks on the MatBench band-gap benchmark. An autonomous coding agent achieves the most accurate model trained without external pretraining, illustrating the potential of autonomous AI loops for materials discovery.
We rethink generative reconstruction attacks against Graph Neural Networks. We analyze vulnerabilities and propose two novel graph inversion attacks to recover training graphs or sensitive structures from model outputs, highlighting privacy risks of GNNs.
We propose Never Skip a Batch, a method for dense learning of Temporal GNNs via Adaptive Pseudo-Supervision. Moving-Averaged Labels provide soft targets from past supervision to fill gaps when labels are sparse, improving gradient stability and convergence without changing the model or loss.
We propose Freeze, Prompt, and Adapt, a framework for source-free unsupervised GNN prompting. The pre-trained GNN is kept frozen and target-domain labels are unavailable; prompts are learned to adapt without labeled data, enabling parameter-efficient transfer.
We present DiRecGNN, an attention-enhanced entity recommendation framework for monitoring cloud services. By constructing a production-scale heterogeneous monitor graph and applying attention over entities, we identify which attributes should be tracked by an automated watchdog, with deployment insights.
We establish a fundamental expressivity limit for Message-Passing GNNs. By defining an information-complexity for aggregations, we prove MP-GNNs with common aggregations yield only polynomially many graph equivalence classes, while the space of graphs is super-exponential; already two iterations of Color Refinement yield exponential class counts.
We introduce DRIFT, a benchmark for task-free continual graph learning with continuous distribution shifts. DRIFT simulates non-stationary graphs without task boundaries and provides evaluation protocols and baselines to study forgetting and continual adaptation.
We propose Federated Graph Learning for EV charging demand forecasting with personalization against cyberattacks. A federated GNN framework leverages spatial correlations across charging stations while preserving privacy and increasing resilience to cyber threats, with personalized models across stations.
We propose generation of uncertainty-aware high-level spatial concepts in factorized 3D scene graphs via Graph Neural Networks. The method discovers concepts such as rooms and walls within factorized scene graphs to support robust navigation and mapping, while accounting for uncertainty in concept discovery.
We propose Alternative Graph Neural Networks, synergizing generalized extreme value models and deep learning for travel mode choice modeling. The approach uses an alternative graph where nodes are choices and edges encode dependence among alternatives, enabling explicit representation of cross-choice correlations.
We propose RelBall, a relation ball with quaternion rotation for knowledge graph completion. RelBall models relations with quaternion rotations to capture non-commutative composition and other relational patterns, improving knowledge graph completion performance.
We propose Ontology-Guided Evidence Path Inference for multi-hop KGQA (OPI). OPI uses a relation-centric ontology graph to constrain evidence paths, reducing search space and ensuring semantic constraints are satisfied in multi-hop questions.
GRAFT introduces benchmarks for linking gene expression to phenotypes in Arabidopsis thaliana, addressing the genome-to-phenome (G2P) challenge. The work emphasizes reasoning over high-dimensional, heterogeneous biological data and motivates the need for graph- and hypergraph-based datasets to evaluate G2P methods in plant genetics and breeding.
This work benchmarks multi-modal graph-based approaches for social media popularity prediction. It notes that accuracy is hampered by not jointly modeling multimodal content and temporal social interactions, and proposes standardized benchmarks across datasets, modalities, and observation windows to enable fair comparisons and progress.
We formulate frame-level freeway risk assessment as a multi-agent scene-graph binary classification. Construct per-frame graphs 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. Introduce hia-gat, a heterogeneous interaction-aware graph attention network, built on this benchmark.
KG2Cypher presents a data-centric pipeline for building enterprise text-to-Cypher systems from existing KGs. It first constructs executable Cypher queries from observed graph facts, then uses LLMs to generate corresponding natural-language questions, and validates results with an LLM judge and human validation.
GNBAN addresses long-horizon forecasting over large entity sets using Graph Neural Basis Attention Networks. It targets tens of thousands of correlated time series across products, stores, and regions, aiming to scale, capture shared dynamics, and remain interpretable, comparing to classical per-series models and other graph-based forecasters.
Beyond Sparse Supervision: Diffusion-Guided Learning for Few-Shot Graph Fraud Detection. The paper tackles sparse and imbalanced supervision and representation dilution in fraud graphs, introducing diffusion-guided learning to improve detection under few labeled fraud cases.
Hybrid Fact-Checking that Integrates Knowledge Graphs, Large Language Models, and Search-Based Retrieval Agents Improves Interpretable Claim Verification. We present a three-step autonomous pipeline: knowledge graph retrieval for rapid one-hop evidence, LLM-based reasoning to generate answers, and an explicit validation step using evidence.
Pulmonary Embolism Risk Stratification from CTPA and Medical Records: Vascular Graphs Are Not All You Need. The study benchmarks models that classify PE risk using either medical records and biomarkers extracted from CTPA images, evaluating whether full vascular graph information is essential.
Attribution Bias in Philosophical Knowledge Graphs: Corpus Frequency versus Temporal Sourcing. The work argues that corpus frequency drives traditional attribution in KGs, conflating textual power with historical priority, and shows that many top concepts precede their attributed traditions by centuries to millennia.
Travel-Oriented Reasoning Large Language Model via Domain-Specific Knowledge Graphs. The paper proposes a modular pipeline grounding travel-domain reasoning in a domain graph to improve accuracy and reliability, addressing model hallucinations in specialized domains.
HyperSU: Corpus-Driven Semantic-Unit Hypergraph for Retrieval-Augmented Generation. It critiques HyperRAG reliance on LLM-generated summaries to form hyperedges due to hallucinations and indexing costs, and proposes corpus-driven semantic-unit hypergraphs with alternative retrieval strategies beyond one-hop or PageRank diffusion.
LLM based Knowledge Graph Approach to Automating Medical Device Regulatory Compliance. The work discusses automating parsing and cross-referencing CFR Title 21 regulations using KG and LLMs to reduce compliance burden for medical devices.
Beyond the Reranker: Do RAG Retrieval Enhancements Help Once a Strong Reranker Is Present? The study evaluates various retrieval enhancements (query expansion, hierarchical/cross-document summarization, graph-based expansion, etc.) in mixed-format corpora and finds that benefits are limited when a strong reranker is already present.
SemFlowRAG: Directed Semantic Flow from Abstraction to Evidence for Complex Reasoning. It reconstructs the retrieval space into a corpus-adaptive semantic graph to prevent probability black holes and semantic drift, enabling directed information flow from abstract concepts to evidence.
Multimodal Graph RAG for Long-range Visually Rich Document Understanding. The method uses a document-level knowledge graph to summarize global document knowledge and support holistic understanding for long-range VQA, addressing limitations of page-level retrieval.
Efficient Retrieval-Augmented Generation via Token Co-occurrence Graphs. TIGRAG builds a token co-occurrence graph to enable efficient, scalable RAG without heavy LLM-based extraction pipelines; it reduces computational burden while maintaining reasoning capability.
Ontology-Compliant Knowledge Graphs. The paper discusses building KGs that comply with ontology schemas, offering term-matching algorithms, pattern-based compliance approaches, and metrics, with a building-sector case study.
RankGraph-2: Lifecycle Co-Design for Billion-Node Graph Learning in Recommendation. It presents Meta-scale graph retrieval by co-designing graph construction, representation learning, and real-time serving, using a co-learned cluster index to avoid online KNN costs.
Latent Bridges for Multi-Table Question Answering. GRAB introduces a constructor-encoder-bridge pipeline that lifts relational data into a heterogeneous graph, encodes via message passing, and passes a small set of query-conditioned latent tokens to a frozen LLM, training only the graph encoder and latent bridge.