Showing 24 papers for 2026-05-25
This work presents a scalable heterogeneous graph neural network workflow for data-driven optimal power flow (OPF) in smart grids. Built on HydraGNN, it preserves node and edge type heterogeneity to train surrogates and develop graph foundation models that generalize across grid topologies.
We propose a self-supervised adversarial purification framework for GNNs that separates robustness from the classifier by inserting a dedicated purifier that cleans inputs before classification. Unlike previous purification methods, purification is learned with self-supervision, enabling improved robustness without sacrificing accuracy.
RelPrism is a multi-faceted pre-training framework for relational databases that converts RDBs into graphs and uses self-generated tasks to learn representations capturing cross-table interactions. The pre-training supports diverse predictive tasks on relational data and improves downstream performance.
We prove that for every k, k-WL cannot distinguish all non-isomorphic simple-spectrum graphs, showing incompleteness of WL-based GNNs on this class. To close the gap, we propose PRiSM, a complete canonicalization procedure (Partition, Refine, Solve, Match) that enables provable isomorphism testing.
S^3GNN enables efficient long-range graph learning by combining global spectral mixing with local message passing to mitigate oversquashing. It leverages spectral operators to improve information propagation across distant nodes while stabilizing propagation dynamics.
This work enhances GNN explainability by introducing Relevant Walk Search, which identifies informative walks as explanations with higher-order information flows. It improves upon GNN-LRP by reducing the computational cost from exponential to tractable levels.
The paper introduces learning stability landscapes for synchronization networks, a graph-to-image prediction task where per-node targets are image-like landscapes derived from topology. This approach provides deeper insights into network robustness and enables deriving multiple scalar stability indices.
SeedER is a retrieval framework for knowledge graphs that seeds a compact candidate set and iteratively expands via low-cost exploration, enabling efficient multi-hop query retrieval on large-scale KGs. It explicitly leverages KG structure to reduce computational costs.
A cross-attention-based bipartite GNN is proposed to couple nodal fields and elemental fields in large-deformation sheet forming, enabling direct node-element interaction modeling rather than post-hoc interpolation. This accelerates finite element simulations by better capturing node-element coupling.
The paper studies the expressive power of Deep Homomorphism Networks (DHNs) over relational databases, showing their alignment with SQL fragments and first-order logic. Through analysis of max, sum, and mean aggregations, it clarifies the representational capabilities of DHNs.
SciAtlas introduces a large-scale knowledge graph for automated scientific research to support topological reasoning beyond keyword search. It enables agentic, reasoning-based exploration while aiming to mitigate hallucinations.
Heterogeneous Sheaf Neural Networks (HetSheaf) propose a unified framework for learning on heterogeneous graphs using sheaf theory. By reducing reliance on specialized modules, HetSheaf enables principled integration across node and edge types.
Graph Learning via Logic-Based Weisfeiler-Leman Variants and Tabularization presents a pipeline that tabularizes graphs using WL variants guided by logic. It characterizes expressive power with a generalized bisimulation game and reports results on 14 datasets.
GILT is an LLM-free, tuning-free graph foundational model designed for in-context learning on graphs. It addresses extreme graph heterogeneity without relying on large language models or fine-tuning.
Bridging Data and Physics: A Graph Neural Network-Based Hybrid Twin Framework proposes a GNN-based hybrid twin that combines data-driven models with physics-based simulations (e.g., FEM). It aims to reduce the ignorance gap between data and physics and improve predictive accuracy in unobserved scenarios.
GP2F advances cross-domain graph prompting by adaptively fusing pre-trained GNNs, enabling robust transfer to downstream tasks with domain shifts. It explains why graph prompting remains effective under cross-domain distribution changes.
MemReward introduces a graph-based experience memory for LLM reward prediction under limited labels. It enables better reward estimation when ground-truth annotations are scarce by leveraging relational memory graphs.
Interactive Query Answering on Knowledge Graphs with Soft Entity Constraints tackles queries with vague or context-dependent constraints. It proposes interactive, constraint-tolerant query answering to better reflect user preferences.
Graph Alignment Topology as an Inductive Bias for Grounding Detection proposes using graph alignment topology as a prior to improve grounding detection in LLM outputs. This inductive bias encodes topological cues to enhance factual grounding.
MUSEKG presents MuseKG, an interactive knowledge graph over museum collections that organizes heterogeneous data into a typed graph linking objects, people, organizations, and images. It enables relation-aware exploration and integrated navigation of museum data.
PLACE (Prompt Learning for Attributed Community Search) introduces a graph prompt learning framework for attributed community search in large graphs. Inspired by NLP prompt-tuning, it inserts learnable prompt tokens into the graph to contextually refine queries, creating a prompt-augmented graph where the prompts bridge the structural information and the query, enabling query-dependent refinement that improves retrieval of target communities.
KG-R1 is an agentic framework that optimizes knowledge-graph retrieval-augmented generation (KG-RAG) via reinforcement learning. Unlike traditional modular pipelines of LLM components (planning, reasoning, responding), it learns a cohesive policy to coordinate retrieval, reasoning, and generation, reducing inference costs and improving transferability across different KG schemas while keeping verifiable reasoning traces grounded in the knowledge graph.
Concomitant DAG Learning investigates how noise adaptivity, sparsity, and non-negativity constraints affect learning DAG structures from observational data. The work addresses the combinatorial challenge of enforcing acyclicity and scalability in causal discovery, proposing a methodology that leverages these aspects to improve identifiability and robust structure recovery.
STEC-Net is a spatiotemporal graph neural framework for community discovery in dynamic social networks. By jointly modeling spatial structure and temporal evolution, STEC-Net produces embeddings that capture both where communities are located and how they evolve over time, enabling more accurate and robust community detection in dynamic networks.