Showing 13 papers for 2026-03-30
This paper extends KGWAS by incorporating contextual information into the knowledge graph workflow to improve GWAS interpretability and discovery power. By conditioning on context-specific biology rather than a single large general KG, it aims to yield more mechanistic, testable variant–gene–interaction findings.
We propose a framework to assess the robustness of LLM-enhanced GNNs under poisoning attacks that manipulate both graph topology and textual features. The framework enables systematic evaluation and reveals how semantic features from LLMs affect vulnerability, guiding more robust designs.
DPD-Cancer is a graph-based deep learning pipeline for predicting small-molecule anti-cancer activity across cancer cell lines, with explanation capabilities. It integrates chemical structure and cellular context to capture non-linear relationships and heterogeneity, delivering interpretable predictions.
PEANUT introduces a gradient-free, restricted black-box attack that perturbs GNNs by aligning eigenvalues and injecting virtual nodes to exploit topology-driven message passing. The method highlights a simple yet effective vulnerability of adjacency/Laplacian-based GNNs under topology perturbations.
This work analyzes the computational complexity of optimally rewiring graph topology to mitigate oversmoothing and oversquashing in deep GNNs. It provides hardness results and theoretical limits, informing how feasible topology optimization is at scale.
Geometric Evolution Graph Convolutional Networks (GEGCN) model geometric evolution on graphs using discrete Ricci flow, with an LSTM capturing the evolution and feeding dynamic representations into a GCN. Experiments show state-of-the-art performance on multiple classification benchmarks.
We propose a topology-aware RL framework for optimal dispatch of energy storage systems in distribution networks, using TD3 and GNN-based graph feature encoders to adapt to topology changes. The study systematically compares three GNN variants and demonstrates improvements in economic performance and voltage security.
D-GATNet is an interpretable temporal graph attention network for ADHD identification from dynamic functional connectivity derived from fMRI. It models time-varying brain networks with attention mechanisms to improve classification and provide interpretable insights.
This work combines Vision Transformers and Graph Neural Networks to improve charged particle tracking in the ATLAS Muon Spectrometer under high-luminosity conditions. The approach aims to robustly identify and reconstruct muons with higher efficiency in crowded environments.
Causal Graph Neural Networks for Healthcare discusses combining graph-based representations with causal inference to learn invariant, cross-institutional mechanisms, addressing brittleness and biases in real-world deployments.
The paper assesses reproducibility and artifact consistency of SIGIR 2022 graph-based recommender system papers, identifying gaps in data, code, evaluation, and reporting. It offers recommendations to improve methodological rigor and replicability.
ReCUBE introduces a benchmark to measure how effectively LLMs leverage repository-level context for code generation, by tasking them to reconstruct a masked file using remaining code and project context. It reveals variability in contextual utilization across models.
The study argues that simply exposing private-library API docs is insufficient for accurate private-library code generation. It explores methods to teach LLMs to effectively use private libraries for code generation, focusing on instruction and training strategies that enhance API invocation.