Showing 23 papers for 2026-04-23
This paper studies how graph neural networks can be used on edge intelligent meters in a microgrid to forecast PV power. It describes background technologies including ONNX and ONNX Runtime, and outlines the meter hardware/software. It then trains and deploys two graph neural network models, GCN and GraphSAGE, with a focus on developing and deploying a customized ONNX operator for GCN.
This study evaluates the baseline D(G)-ζ(G) polynomial model for molecular measurements on MoleculeNet datasets. Graph-theoretic indices offer interpretability and low computational cost, but generalization to larger chemically diverse datasets is uncertain; the paper tests whether the D(G) and ζ(G) model generalizes beyond a small flavonoid dataset.
This work studies learning over the graph Laplacian L in a distributed streaming setting, where new edges arrive in real time across a network of workers. Maintaining a consistent distributed representation of L while learning is hard; the authors propose ridge spectral sparsification to produce a sparse surrogate that preserves spectral information, enabling faster, scalable learning on evolving graphs.
Concept Graph Convolutions introduce a graph convolution designed to operate on node-level concepts, enhancing interpretability of message passing. The layer performs message passing in the concept space, providing more transparent reasoning about predictions.
Cross-sectional stock ranking suffers from crosstalk between predictive factors. The paper identifies two forms: temporal-scale crosstalk and inter-stock information interference. They propose Anti-Crosstalk Learning (ACT) to disentangle temporal dynamics and purify inter-stock structure, improving ranking accuracy.
Fault location in partially observable distribution grids, challenges; investigates spatio-temporal GNNs; evaluate topology definitions beyond full-grid.
StormNet is a spatio-temporal graph neural network designed for bias correction of storm surge forecasts, addressing uncertainties in traditional numerical models like ADCIRC. It integrates bias correction within a GNN framework to improve predictive accuracy across coastal regions. The approach combines physical reasoning with data-driven learning for robust forecasts.
F2LP-AP proposes a fast, training-free label propagation method with an adaptive propagation kernel for semi-supervised node classification. It addresses computational overhead of GNNs and heterophily by replacing training with an adaptive, efficient propagation process. Empirical results show strong performance and scalability.
This paper presents a proof-of-concept Graph Neural Network model that can successfully predict network flow-level NetFlow by accurately modelling the graph structure and the connection features. We use sliding-windows to split the network traffic into equal-sized heterogeneous bidirectional graphs containing IP, Port, and Connection nodes. Then we use the GNN to model the evolution of the graph structure and the connection features. Our approach shows superior results when identifying the Port and IP to which connections attach.
This work introduces a gauge-equivariant graph neural network that enforces local non-Abelian gauge symmetry within message passing via matrix-valued, gauge-covariant features and symmetry-consistent updates. It extends equivariant learning from global to fully local symmetries, enabling principled learning for lattice gauge theory observables, including nonlocal quantities. Experiments demonstrate improved physics-consistent predictions.
From Nodes to Narratives: Explaining Graph Neural Networks with LLMs and Graph Context. GSPELL, a lightweight, post-hoc explanation approach for GNNs on text-attributed graphs. It uses LLMs and graph context to produce interpretable rationales.
This paper proposes R2RAG-Flood, a training-free retrieval-augmented generation framework for flood damage nowcasting with reinforced reasoning. It builds a knowledge base from labeled tabular records, where each sample includes structured predictors, a compact text-mode summary, and a model-generated reasoning trajectory. During inference, the target prompt is augmented with geographically local neighbors and selected free-shots to support case-based reasoning without task-specific fine-tuning. A two-stage procedure retrieves relevant records and then generates the final answer.
Colorful Talks with Graphs: Human-Interpretable Graph Encodings for Large Language Models. Investigates how LLMs can be applied to graph problems by introducing human-interpretable structural encodings for graph-to-text translation. The encoding preserves graph structure and invariances, enabling LLMs to reason about graphs more effectively.
LLM Agents Grounded in Self-Reports Enable General-Purpose Simulation of Individuals. Using data from a national sample of 1,052 Americans, the authors build person-specific simulations by grounding LLM agents in self-report data. They show that such agents can simulate individual behavior across domains, enabling generic, controllable generative simulations.
Unlocking the Edge deployment and on-device acceleration of multi-LoRA enabled one-for-all foundational LLM. Deploying LLMs on smartphones with hardware-aware acceleration and multiple LoRA adapters, integrated into a single frozen inference graph for on-device multilingual use cases on Galaxy devices.
Choose Your Own Adventure: Non-Linear AI-Assisted Programming with EvoGraph. Introduces EvoGraph, an IDE plugin that combines AI interactions and code changes into a lightweight development graph to support non-linear, branching AI-assisted programming. It helps developers explore alternatives, manage prompts, and trace changes.
Location Not Found: Exposing Implicit Local and Global Biases in Multilingual LLMs. LocQA introduces a test set of locale-ambiguous questions across 12 languages to quantify inter- and intra-lingual biases in multilingual LLMs. The study reveals biases in local facts and cross-language consistency.
Revisiting Catastrophic Forgetting in Continual Knowledge Graph Embedding. The work challenges the view that continual KGEs forget only by altering old embeddings, showing that new entities can interfere with existing embeddings, degrading previous task performance. It analyzes interference mechanisms and suggests directions to mitigate.
Impact of large language models on peer review opinions from a fine-grained perspective: Evidence from top conference proceedings in AI. The study analyzes how LLMs influence peer review, evaluating linguistic form, originality, and evaluation aspects in top AI conference proceedings.
Reasoning Over Space: Enabling Geographic Reasoning for LLM-Based Generative Next POI Recommendation. ROS introduces geography as a vital decision variable within LLM reasoning by a Hierarchical Spatial Semantic ID that discretizes locality and POI semantics into tokens, enabling geographic-aware recommendations.
This paper investigates a remote Rowhammer-style vulnerability in Federated Learning (FL) systems, showing how adversarial observations from distributed FL clients can induce memory faults at the central server. It argues that FL’s reliance on sparse gradient updates and remote direct memory access creates a new security surface beyond data privacy, where clients could meaningfully disturb the server, and discusses potential defenses against such attacks.
The authors address the generalization challenges of applying machine learning to AC Optimal Power Flow (ACOPF) by introducing a Hybrid Heterogeneous Message Passing Neural Network (HH-MPNN). This architecture combines a heterogeneous GNN with a scalable transformer and physics-informed positional encodings to better handle varying grid topologies. The model aims to improve both scalability and topology-flexible generalization while maintaining solution quality for large-scale power systems.
This work tackles covariate distribution shift in graph data, noting that many real-world graphs exhibit OOD characteristics that standard GNNs fail to handle. It proposes graph data augmentation guided by contrastive learning to leverage latent information, aiming to better capture covariate shifts and improve robustness of GNNs under distributional changes.