Showing 18 papers for 2026-05-07
GraphPI treats protein inference as a node classification problem on a protein-peptide-PSM graph. It leverages graph neural networks to propagate information across interconnected proteins through the peptide and PSM connections, addressing data scarcity in protein annotation. The approach aims for efficient and accurate protein inference leveraging network structure.
HeterSEED introduces semantics-structure decoupling for heterogeneous graphs under heterophily. Traditional metapath-based aggregation can mislead when semantic roles diverge from feature similarity. The framework decouples relational semantics from structural propagation to improve robustness and performance.
We propose polarity-aware representation learning on clause-literal hypergraphs for unsat-core prediction in SAT. By modeling higher-order interactions among literals and clauses beyond bipartite graphs and incorporating literal polarity, the method improves unsat-core prediction accuracy.
ALL-IN enables transferability across graphs with different input feature spaces by projecting node features into a shared random space. The approach provides a theoretically grounded way to construct dataset-agnostic representations, enabling more robust graph foundation modeling.
Quantile-free Prediction Interval GNN (QpiGNN) offers uncertainty quantification for graphs without relying on resampling or post-hoc calibration. It builds on quantile regression concepts to directly optimize covariate-related prediction intervals, addressing non-exchangeability concerns in graph data.
We model surgical team dynamics in real time using time-expanded interaction graphs, where team members are time-indexed nodes and communication exchanges form directed edges. This structured representation enables actionable insights for intraoperative coordination and decision support.
Joint relational database generation is achieved via graph-conditional diffusion models, enabling parallel, coherent generation of multi-table relational databases. This approach overcomes the bottlenecks of autoregressive generation and improves downstream applicability for privacy-preserving data release and augmentation.
DPD-Cancer is a graph-attention deep learning framework for predicting small-molecule anti-cancer activity across the NCI-60 panel, using chemistry-aware data partitioning. It achieves AUROC 0.87 and AUPRC 0.73 on hold-out data and provides explainable insights into predictions.
Geometric Evolution GCN (GEGCN) enhances graph representation learning by modeling geometric evolution via discrete Ricci flow with an LSTM, feeding dynamic representations into a graph convolutional network. It shows strong performance on various classification tasks.
CausalGaze introduces counterfactual graph interventions to unveil hallucinations in large language models. Moving beyond static signals, it uses active causal interventions on graphs to detect and interpret when LLM outputs are hallucinatory.
Learning time-varying graphs from incomplete graph signals presents a unified non-convex optimization framework that jointly recovers a sequence of graph Laplacians and imputes missing signal entries. This bidirectional graph-signal coupling improves robustness under high missingness.
A large language model–type architecture for high-dimensional molecular potential energy surfaces represents molecular systems as graphs with nodes, edges, and faces, and uses LLM-inspired generative mechanisms to model interactions and generate PES.
NoisyCausal provides a benchmark for evaluating causal reasoning under structured noise, highlighting how observations corrupted or irrelevant can impair causal inference and offering evaluation protocols for robust reasoning.
GEM combines graph-structured dialogue understanding with ReAct agents and a mixture-of-experts framework to enhance dialogue state tracking, dynamically routing between specialized experts including a graph neural network.
CodeEvolve extends LLM-driven optimization with runtime-enriched target selection, Monte Carlo Tree Search, automated code refinement, and language-specific evaluation pipelines to optimize multi-language code (Java and Apex) efficiently.
From Beats to Breaches analyzes how Offensive AI can infer sensitive user information from public playlists, introducing musicPIIrate to quantify privacy risks and demonstrate potential leaks.
KGLAMP proposes a knowledge graph–guided LLM planning framework for adaptive planning and replanning in heterogeneous multi-robot systems, addressing heterogeneity and uncertainty without manual symbolic models.
A Graph-Native Approach to Normalization discusses normalization of knowledge graphs, highlighting issues from schema-less LPGs and proposing a graph-native normalization that leverages node-level dependencies to reduce redundancies and inconsistencies.