Showing 19 papers for 2026-08-10
MiGHT-EHR introduces a multi-task graph transformer designed for heterogeneous temporal EHRs. It models multiple entity types (patients, visits, diagnoses, prescriptions, procedures) and their temporal interactions to enable joint predictions across tasks. It captures longitudinal trajectories and cross-entity dependencies to improve clinical prediction.
We present SNI-GNN, a SmartNIC-assisted full-graph training system for GNNs. It reduces inter-node embedding communication by predicting remote embeddings in-network using a lightweight predictor on SmartNICs, and couples an importance-based boundary-node sampling with an asynchronous DPU–GPU data pipeline and intermediate-result reuse. The approach preserves accuracy while achieving better scalability on multi-server clusters.
Accounting Graph Transformer (AGT) enables joint 12-month forecasting of 13 KPIs from 71 monthly ledger series in small businesses with short histories. Each ledger series is represented as a masked token and information is exchanged via a fixed accounting-relations graph using typed attention, with target-specific context pooled for accurate forecasts.
TRicci provides edge sparsification for dynamic graphs using Temporal Forman-Ricci curvature. It extends Forman-Ricci curvature to dynamic networks to guide which edges to drop while preserving essential structure for learning. This reduces computation in dense, rapidly evolving graphs without severely sacrificing performance.
The paper analyzes GNN-based forecasting where time-series are nodes and edges are pairwise temporal correlations, comparing static graph assumptions versus dynamic correlations. It quantifies temporal correlation volatility and its impact on forecasting accuracy and proposes strategies to adapt GNNs to evolving graphs, improving robustness.
The paper introduces a joint factor-graph framework that uses locality by exposures to diversify portfolios. It combines traditional factor models with graph-based market structure, enabling allocations to be driven by underlying exposure networks rather than solely observed co-movements. This approach reduces vulnerability to idiosyncratic shocks.
MolBioKG tackles out-of-graph molecules by grounding unseen molecules to biomedical evidence via multi-resolution anchoring. It connects an index of 2.74 million molecules (scaffolds, fragments, functional groups, fingerprints) to a 9.6-million-edge knowledge graph.
TMTE addresses multimodal attributed graphs with topology quality issues, proposing co-evolution of task-aware modalities and topology. It adapts modality contributions and graph structure in a task-driven manner to handle noisy or missing relations and improve downstream performance.
The paper introduces cluster attention to enlarge the effective receptive field without full global attention by partitioning the graph into clusters and applying attention within clusters. This yields scalable graph transformers with graph-structure biases and reduced computation, while preserving long-range dependencies.
The work studies combining LLM agents with knowledge graphs to predict urban socioeconomic indicators. LLMs provide reasoning and task-relevant knowledge while KGs supply structured data, reducing reliance on handcrafted features and improving predictive accuracy.
BDDs are used in reversible circuit synthesis; variable ordering is NP-hard. BDD2Seq uses graph-to-sequence learning to predict variable orderings, enabling scalable synthesis and more efficient quantum circuit design.
Integrates fragment-based GNNs with many-body expansion to predict PES for large molecular systems. It introduces data-adaptive transfer learning to transfer knowledge across molecules, reproducing first-principles PES for large systems.
The paper proposes multi-label graph foundation models that enable transfer across graphs by moving beyond single-vector representations to multi-semantic basis learning. This enables cross-domain generalization for multi-label node classification.
The paper proposes DKG-MTI, a dual knowledge graph framework for unified multi-task user intent inference from online travel reviews. It builds a User-Specific Intent KG and aligns it with a Global Hotel KG to improve inference while reducing error propagation.
ReGraph learns to generate recipe graphs from food images to capture structured procedural knowledge. By modeling ingredients, states, and actions as a graph, it provides a framework to assess whether models encode process-level knowledge in recipe generation.
INSPECT-AI is an LLM-based interactive tool that assists human reviewers with research integrity assessments of RCT publications. It uses provenance knowledge graphs to document decisions, enabling transparency and reproducibility in systematic reviews.
The paper proposes a graph-based model that predicts single-cell transcriptional responses to unseen perturbations, anchored by gene geometry. It leverages network priors to avoid overfitting to perturbation-responsive pathways and emphasizes stable gene relationships in modeling.
The paper proposes a hybrid KG generation pipeline that grounds an LLM in Wikidata with multilingual coverage and uses an agentic reflexion pattern to synthesize emerging concepts and metadata. It balances top-down grounding with bottom-up expansion to build scalable, multilingual knowledge graphs.
We present a physics-inspired digital twin for brain activity using a band-stratified metriplectic port-Hamiltonian neural network trained on EEG data. The port-Hamiltonian formulation enforces passivity and a power balance, offering robust modeling for BCI applications.