Showing 27 papers for 2026-02-17
We propose a federated learning approach to learn nonlinear temporal dynamics across heterogeneous clients without sharing raw data. Graph attention is used to provide cross-client interpretability of inter-subsystem relations.
We introduce an optimization-free graph embedding method based on a distributional kernel to tackle over-smoothing in NAS/WL. By leveraging node and degree distributions, it yields expressive representations for community detection without iterative optimization.
GREPO provides a benchmark to evaluate GNNs on repository-level bug localization where processing entire repositories exceeds typical LLM context windows. It compares graph-based methods to retrieval and simple heuristics, highlighting the benefits of structural code representations.
OPBench introduces a comprehensive graph benchmark for evaluating graph learning methods across real-world opioid crisis scenarios. It standardizes datasets and tasks to enable systematic comparisons and analysis.
We study instance-dependent algorithm selection for combinatorial auctions by training GNNs to predict instance hardness for greedy allocation. This enables choosing solvers dynamically and improves performance on hard instances.
We propose BHyGNN+ for unsupervised representation learning on heterophilic hypergraphs. It extends prior BHyGNN to work without labels, improving representations when connected nodes belong to different classes.
GNNs trained to predict observable dynamics can decompose complex neural activity into simple, interpretable representations. They reveal connectivity, neuron types, signaling functions, and sometimes hidden external stimuli in simulated neural assemblies.
NeuroMambaLLM combines dynamic graph learning of fMRI functional connectivity with Mamba state-space reasoning and language-model-based interpretation. It captures transient neural dynamics in autistic brains and supports reasoning-driven insights.
We propose a hybrid TGN-SEAL model for dynamic link prediction that blends Temporal Graph Networks with SEAL-inspired neighborhood features. It addresses sparsity and class imbalance in evolving networks and improves predictive performance.
GREAT-EER introduces graph edge attention for emergency evacuation planning under the Bus Evacuation Orienteering Problem. It models evacuation decisions over a graph and demonstrates effectiveness under NP-hard BEOP constraints.
Return of the Schema provides workflows and datasets to bring complete schema-level knowledge into ML and reasoning on knowledge graphs. It enables evaluation of ontological constraints and neurosymbolic methods on large-scale KGs.
GraphFM is a generalist graph transformer pretrained on diverse graphs with a Perceiver-based encoder and latent tokens to learn transferable representations. This design aims to enable cross-domain transfer and reduce dataset-specific tuning.
We show that heuristic methods can serve as teachers to distill MLPs for graph link prediction. Beyond standard GNN teachers, we include GNN4LP and heuristics such as common neighbors as teachers, enabling effective distillation.
HYPER is a foundation model for inductive link prediction on knowledge hypergraphs, capable of generalizing to unseen relation types. It enables scalable reasoning on hypergraphs by leveraging pretraining across diverse tasks.
We explore three levels of modeling for delay prediction: a heterogeneous GNN baseline with attention, FlowKANet using Kolmogorov-Arnold Networks to replace MLPs, and KAMP-Attn for attention. FlowKANet reduces trainable parameters while maintaining competitive performance.
We introduce cardinality-preserving attention channels for graph transformers in molecular property prediction. CPA channel preserves dynamic signal related to support size, complementing static embeddings, and is coupled with structured sparse attention and self-supervised pretraining.
RPG-AE combines a Graph Autoencoder with rare pattern mining for provenance-based anomaly detection. It builds a process behavioral graph via k-NN and learns the normal relational structure to detect APT-like activities.
We enable robots to discover high-level spatial concepts from primitive observations within factorized 3D scene graphs, while accounting for uncertainty. These concepts are incorporated as optimization factors to constrain relative geometry, improving graph-SLAM performance.
We apply graph neural networks to interferometer simulations, capturing optical physics behind LIGO-scale problems. GNNs achieve runtimes 8-15x faster than state-of-the-art solvers while maintaining accuracy.
We present LLM-Enhanced Rumor Detection via Virtual Node Induced Edge Prediction. The framework uses LLMs to model textual coherence along rumor propagation and adds virtual nodes to induce edges for improved detection.
GraFSTNet proposes a unified framework that combines graph-based spatial modeling with frequency-domain temporal modeling to capture both spatio-temporal dependencies and periodic patterns in cellular traffic, enabling accurate predictions. It addresses the limitations of prior methods that focus only on temporal modeling or rely on predefined topologies by jointly modeling spatio-temporal dynamics.
This work studies large-scale brick kiln detection using high-resolution satellite imagery, creating a multi-city dataset with over 1.3 million image tiles across five regions. It introduces ClimateGraph, a region-adaptive graph-based model that captures spatial and directional layouts of kilns, and demonstrates its effectiveness for scalable detection compared to baselines.
G2CP introduces a graph-grounded communication protocol for multi-agent reasoning with LLMs, using graph operations over a shared knowledge graph instead of free-text messages. Agents exchange traversal commands, subgraph fragments, and updates, enabling verifiable reasoning traces and reducing semantic drift and inefficiency.
This work proposes anonymization-enhanced privacy protection for mobile GUI agents, enabling automation while safeguarding sensitive screen content. The framework performs automatic anonymization to protect privacy without sacrificing task-relevant information, making mobile GUI automation more privacy-friendly.
Agentic Assistant for 6G presents turn-based conversations to enable hierarchical co-management of AI-enabled RAN and edge AI. It addresses real-time control and expertise scarcity, moving beyond basic retrieval-augmented generation to structured, turn-based interactions for coordinated network management.
Rank-GRPO introduces a two-stage ConvRec-R1 framework to train LLM-based conversational recommender systems with reinforcement learning, aiming to improve ranking quality and item validity. Stage 1 prepares data and alignment, while Stage 2 applies RL fine-tuning with ranking objectives to prevent out-of-catalog items and formatting errors.
PersonalAI systematically compares knowledge-graph storage and retrieval approaches for personalized LLM agents and proposes an external memory framework based on a knowledge graph that is automatically constructed and updated by the LLM, enabling scalable, structured memory for long-term personalization.