Showing 24 papers for 2026-03-18
This paper proposes a federated learning framework that combines a medical knowledge graph with a temporal transformer and meta-learning to enable privacy-preserving early sepsis prediction across multiple ICUs. It addresses data fragmentation and the temporal complexity of clinical data, aiming to improve predictive accuracy while preserving patient privacy.
The paper develops a Graph Neural Network trained on a watershed connectivity graph to map flash flood susceptibility in Himachal Pradesh using six years of Sentinel-1 flood inventory and 12 environmental variables. It incorporates conformal uncertainty quantification to provide reliable, downstream-aware risk maps rather than pixel-wise predictions.
We show that introducing discrete Ricci flow into hypergraph neural networks can regulate node feature evolution and alleviate over-smoothing in deep HGNNs. Building on this, we propose a diffusion-based model guided by Ricci flow (RFlow) for hypergraphs to improve representation learning.
The paper provides a depth-aware comparison of Euclidean and hyperbolic GNNs on Bitcoin transaction systems, analyzing how neighborhood depth affects predictive performance for tasks such as fraud detection. Using the Elliptic dataset, it evaluates how geometry choices interact with multi-hop contexts.
RaDAR is a relation-aware diffusion-asymmetric graph contrastive learning framework for recommendation. It tackles issues from random edge perturbations that distort signals and data sparsity that limits signal propagation, improving robustness and generalization.
GIST introduces gauge-invariant spectral transformers for scalable graph neural operators. By avoiding expensive exact eigendecomposition and preserving gauge symmetry, it enables scalable, generalizable learning across different spectral decompositions.
ReFORM aggregates user decision factors from reviews into profile generation using a large language model with multi-factor attention for restaurant recommendation. It combines LLM-generated descriptions with review signals and graph-based methods to improve robustness and interpretability.
The work presents an approximate graph-based method to elicit detonation lattices from 3D pressure signals, enabling accurate segmentation of detonation cells. It uses a training-free segmentation approach achieving around 2% error on generated data and yields 3D lattice measurements.
drGT builds a heterogeneous network across drugs, genes, and cell lines to predict drug response with attention-guided interpretability. It evaluates predictive generalization across splits and validates the biological plausibility by comparing to PubMed co-mentions and a structure-based DTI predictor.
LogicXGNN is a post-hoc framework that grounds GNN explanations by constructing logical rules over reliable predicates derived from the model's message-passing. It aims to improve grounding quality for end users while preserving fidelity.
The paper studies merging graph models trained on different domains to form a generalized model. It proposes a graph generation strategy that instantiates a mixture distribution of the domains and merges backbones to create a unified model.
SEPAL is proposed for scalable feature learning on huge knowledge graphs, enabling embedding propagation beyond link prediction with improved memory efficiency. It addresses GPU memory constraints and demonstrates scalable downstream learning.
Controllable graph generation with diffusion models is achieved by inference-time tree search guidance, which steers sampling toward desired properties without retraining. This improves stability and allows incorporating new objectives.
AGRAG advances graph-based retrieval-augmented generation for LLMs by addressing inaccurate graph construction, insufficient reasoning, and incomplete answers. It emphasizes explicit reasoning traces and better grounding to enhance LLM-powered QA.
Enhanced atrial fibrillation prediction in ESUS patients uses hypergraph-based pre-training—both supervised and unsupervised—to improve performance with small cohorts and high-dimensional features. The approach aims for improved accuracy, scalability, and cost-effectiveness.
GNNVerifier proposes a graph-based verifier for LLM task planning to detect and correct potential flaws in plans generated by LLMs. This reduces reliance on unreliable LLM verification and improves plan reliability.
The paper studies probabilistic model selection for GNNs to predict biomedical interactions, focusing on selecting appropriate depth to balance neighborhood information capture. It shows improved predictive performance and mitigates over-smoothing by embracing model selection.
Masked BRep Autoencoder introduces a hierarchical graph transformer for self-supervised learning on CAD BRep models. A masking strategy and a graph-based autoencoder enable reconstruction of masked geometries and attributes from large unlabeled data.
HYQNET develops neural-symbolic logic query answering in non-Euclidean space, decomposing first-order logic queries into relation paths and performing reasoning in hyperbolic space to capture hierarchical structures. It merges neural and symbolic reasoning for improved explainability and performance.
DynaTrust presents a dynamic trust-graph approach to defend multi-agent systems against sleeper agents, which stay benign until triggered. It detects evolving adversaries and reduces false positives by updating trust graphs in real time.
This paper discusses extracting knowledge graphs from biomedical literature to support the ultra-rare disease alkaptonuria (AKU). AKU is caused by mutations in the HGD gene, leading to homogentisic acid (HGA) accumulation and multi-system manifestations; given the scarcity of data, knowledge graphs can integrate scattered clinical and literature information to illuminate disease mechanisms and potential interventions.
IndexRAG introduces a cross-document reasoning approach that moves reasoning from online inference to offline indexing. It identifies bridge entities shared across documents and generates bridging facts as independently retrievable units, enabling cross-document multi-hop reasoning without requiring additional training or fine-tuning.
SentGraph proposes a hierarchical sentence-level graph–based RAG framework to improve multi-hop retrieval augmented QA. Unlike chunk-based methods that produce irrelevant or incoherent context, SentGraph builds a sentence-level graph to connect evidence across documents, yielding more coherent reasoning and better answer quality.
The paper presents three open biomedical knowledge graphs built on Samyama, a high-performance graph database, enabling construction, federation, and AI agent access at scale. It details Pathways KG (118,686 nodes, 834,785 edges from five sources), Clinical Trials KG (7,774,446 nodes, 26,973,997 edges from five sources), and Drug Interactions KG (32,726 nodes, 191,970 edges from three sources).