Showing 16 papers for 2026-06-10
PatchSTG proposes PatchSTG, a patch-based spatiotemporal graph transformer designed to forecast traffic on irregular sensor networks. It tackles uneven sensor distributions and scalability by partitioning the graph into patches and applying efficient attention over local spatiotemporal structures. The method improves forecasting accuracy and efficiency on nonuniform traffic networks.
This work develops a spatiotemporal graph transformer to capture multi-layer interactions in metal additive manufacturing for real-time quality prediction. It leverages rich sensing data to represent repeated layer-wise melting, solidification, and reheating effects across a 3D build, and quantifies contributions of cross-layer interactions to final quality.
We study how benchmark composition affects GNN aggregator choice (sum, mean, max) across 24 datasets. We find edge homophily alone poorly predicts performance gaps; label informativeness predicts GIN-Sum vs GIN-Mean on legacy benchmarks but fails on dense Facebook-100 graphs where labels are nearly uninformative.
We show that evaluating knowledge-graph-completion models with isolated metrics like MRR, Hits@k, and Mean Rank yields conflicting model rankings across datasets. To address this fragmentation, we frame KGC evaluation as a Multi-Criteria Decision-Making problem and propose a meta-analysis approach to aggregate multiple metrics into robust, cross-dataset guidance.
KG-SoftMAP encodes domain knowledge as soft, confidence-weighted, data-overridable priors for Bayesian network structure learning from sparse discrete data. It combines BDeu scoring with a logit-form prior derived from a knowledge graph, enabling better structure discovery when observations are limited.
ERAlign introduces an energy-based representation alignment between GNNs and LLMs on text-attributed graphs to prevent representation drift and improve generalization. It defines alignment objectives that constrain distributions of node representations across models and learns robust cross-modal embeddings.
COGENT is a continuous graph emulator that uses Neural ODEs for long-term physical forecasting on irregular geospatial meshes. A graph-based history encoder builds node-wise context vectors that initialize and condition a latent ODE whose dynamics are driven by interpolated forcing fields.
We introduce dualGNN, an autoregressive GNN for sampling fine, regular triangulations of convex polytopes. The model operates on a dual graph with edges labeled by signed circuits, which capture necessary and sufficient conditions for regularity, and is invariant to the polytope's orientation-preserving symmetries.
We propose a coupled LSTM-GNN framework to reconstruct non-linear, history-dependent local stress fields in heterogeneous microstructures. An LSTM encodes macroscopic stress-strain sequences into a hidden state that captures the constitutive history, while a physics-informed GNN reconstructs the spatial stress field conditioned on that state.
Boosting Graph Robustness Against Backdoor Attacks: An Over-Similarity Perspective argues for robustness enhancements by exploiting over-similarity among benign nodes to obscure triggers, reducing false positives and preserving clean-node accuracy.
Lost in Serialization investigates invariance and generalization of LLM-based graph reasoners under graph serialization changes. It analyzes sensitivity to node labeling, edge encoding, and formatting, studies fine-tuning effects, and proposes principled ways to improve robustness and generalization.
Spatio-Temporal Attention Graph Neural Network (STA-GNN) is proposed for unsupervised anomaly detection in industrial control systems with explainable causal insights, using spatio-temporal attention to reveal causal patterns.
We present a pipeline mapping student questions to curriculum topics using a few-shot classifier anchored in a GPT-4 extracted prerequisite knowledge graph. The approach enables diagnosing knowledge gaps from conversational AI interactions, achieving 80% accuracy across 43 labels on data from a graduate AI course.
LLM-MapRepair enables coherent spatial memory for LLM agents by rectifying incremental maps. It detects, localizes, and corrects structural inconsistencies to build consistent topological graphs as environments grow.
CatalyticMLLM is a graph-text multimodal LLM for catalytic materials, marrying property prediction and inverse structural design in a single framework, addressing representation misalignment between generation and evaluation models.
Graph2Idea introduces Retrieval-Augmented Scientific Idea Generation with Graph-Structured Contexts, enriching ideas with retrieved literature presented as graph contexts to reveal cross-paper relations among problems, methods, mechanisms, and findings.