Showing 15 papers for 2026-04-07
This paper proposes the SmartGuard Energy Intelligence System (SGEIS), an integrated AI framework for electricity theft detection and intelligent energy monitoring in smart grids. It combines supervised machine learning, deep learning-based time-series modeling, Non-Intrusive Load Monitoring NILM, and graph-based learning to capture temporal and spatial dependencies, enabling scalable and reliable detection of non technical losses. It aims to reduce economic losses and improve grid reliability.
Graph transformers promise to overcome oversquashing and long range dependencies, but suffer from quadratic memory and computational costs. The authors explore efficiency-accuracy tradeoffs and introduce k-最大内积注意力, a mechanism that improves efficiency while preserving the expressive power of GraphGPS. This yields scalable graph learning on large graphs and datasets.
This work tackles defect reasoning in Laser Powder Bed Fusion by building a graph-assisted retrieval framework for defect analysis in Ti6Al4V. Scientific publications are transformed into structured representations and relationships among parameters, mechanisms, and defects are encoded into an evidence-linked knowledge graph. This enables agentic reasoning to explain defect formation and guide process optimization.
MAVEN proposes a Mesh-Aware Volumetric Encoding Network to simulate 3D flexible deformation. Traditional GNNs on meshes rely on vertices and edges and miss higher dimensional geometry such as facets and cells. MAVEN uses volumetric mesh-aware encoding to better capture boundary and interior features for accurate deformation prediction and contact handling.
The outage prediction model OPM uses spatially aware hybrid graph neural networks and contrastive learning to forecast electric distribution outages caused by extreme weather. The system provides pre-emptive forecasts to mitigate outage impacts on industry and communities, informing resilience planning and preventative action.
TRACE-KG presents a multimodal framework that goes beyond predefined schemas to construct context-enriched knowledge graphs from complex documents. It jointly builds a context-aware KG while addressing the limitations of ontology-driven pipelines and schema-free extraction, enabling better organization and reuse of long textual information.
This work introduces a graph learning approach for melanoma detection in dermoscopic images based on two representations: superpixel ensemble graphs SEG and superpixel hierarchy graphs SHG. By learning graph structures and weights, the method improves robustness and accuracy of melanoma classification from dermoscopy images.
The paper investigates hallucinations in large language models and proposes a causal graph attention network GCAN to mitigate factual errors. By interpreting internal causal relationships with graph-attention, GCAN aims to improve factual reliability in high-stakes generation such as medical or legal reasoning.
MissNODAG is a differentiable framework for learning cyclic causal graphs from incomplete data, including data missing not at random. It jointly learns the underlying cyclic DAG and the missingness mechanism by integrating an additive noise model with an expectation-maximization procedure.
CATNet applies a geometric deep learning approach using Relational Graph Convolutional Networks to model the CAT bond primary market as a graph. The analysis reveals a scale-free network structure and demonstrates that network-informed spread prediction outperforms baseline models.
Compliance-by-Construction Argument Graphs present a framework to produce evidence-linked formal arguments for certification-grade accountability. The approach structures claims, reasoning, and evidence into argument graphs, while enabling generative AI to assist with drafting explanations and maintaining traceability.
Schema-Aware Planning and Hybrid Knowledge Toolset for Reliable Knowledge Graph Triple Verification proposes a planning-driven framework and hybrid toolset to verify KG triples. It addresses noise in automated KG construction, avoids single-source bias, and improves interpretability of verification results.
GROUNDEDKG-RAG develops a grounded knowledge graph index to support long-document question answering. By grounding retrieval and generation in a KG, it reduces hallucinations, lowers resource consumption, and improves efficiency for long documents.
The Hierarchical Error-Corrective Graph Framework for Autonomous Agents with LLM-Based Action Generation (HECG) integrates multi-dimensional metrics and LLM-driven reasoning to guide action selection. The framework introduces multi-dimensional transferable strategy and an error-corrective graph to align quantitative performance with semantic context, improving action quality.
MisEdu-RAG proposes a misconception-aware dual-hypergraph retrieval augmented generation system to support novice math teachers. It links pedagogical knowledge with student mistakes using a dual-hypergraph RAG to provide actionable instructional feedback.