Showing 15 papers for 2026-04-17
The ST-GAT framework is an explainable GNN-based solution to detect bank distress early and for macro-prudential surveillance in the US interbank network. It models 8,103 FDIC-insured institutions across 58 quarterly snapshots (2010Q1-2024Q2) and reconstructs bilateral exposures from Call Reports into a dynamic directed weighted graph via maximum entropy estimation. The paper emphasizes regulatory-aligned explanations for decisions.
The paper introduces Dual-Path Graph Filtering for fraud detection on graphs, addressing challenges such as relation camouflage, high heterophily, and class imbalance that degrade GNN performance. The method processes messages along two routes to better separate fraudulent vs legitimate patterns and robustly detect fraud.
The paper proposes replacing the Laplacian with a Doubly Stochastic matrix (DSM) in GNNs to capture continuous multi-hop proximity and local centrality. To avoid O(n^3) exact inversion, they approximate DSM with a truncated Neumann series, enabling scalable propagation; DSM serves as the basis for improved message passing.
This work provides a controlled benchmark of node embeddings for graph classification, comparing classical baselines with quantum-oriented representations under a unified pipeline. It evaluates how embedding choices interact with backbones, splits, and training budgets, including circuit-defined variational embeddings and quantum-inspired embeddings via graph operators.
The paper analyzes when GNN-based COVID-19 forecasting outperforms simple baselines, showing that structural sparsification of the input graph and temporal granularity are key factors for performance. By using human mobility networks in Brazil and China, it clarifies conditions under which GNNs offer added value.
The first unsupervised learning model for Maximum Independent Set (MaxIS) in dynamic graphs is presented, combining structural learning from GNNs with a distributed local update mechanism. When edges change, nodes update memories and infer MaxIS membership in a single parallel step; benchmarked against MIP solvers and other methods.
The authors propose a topology-aware active learning framework for graphs, addressing exploration vs exploitation under scarce label budgets. A Balanced Forman Curvature (BFC) based coreset selects representative seeds reflecting cluster structure, with a data-driven stopping criterion and a dynamically triggered shift from exploration to exploitation.
The SDM-SCR framework for graph contrastive learning on text-attributed graphs uses LLMs to decouple semantics into signal and noise views, replacing blind augmentations with semantic-aware perturbations. It then applies Approximate Orthogonal Decomposition to refine representations with structure awareness.
Disentangled dual-branch graph learning for conversational emotion recognition blends dual-space feature disentanglement with dual-branch graph networks. A shared encoder plus modality-specific encoders separate textual, acoustic, and visual cues, while dual graphs model intra- and inter-modal interactions to boost utterance-level emotion accuracy.
This thesis surveys Ollivier-Ricci curvature for metric spaces and graphs, connecting it to classical Ricci curvature and transport-based theory. It discusses extensions by Lin-Lu-Yau on graphs and explores implications for graph neural networks.
Agentic Knowledge Graphs with Recursive Crawling tackle limitations of semantic search in complex enterprise documents. The CFR benchmark shows substantial accuracy gains over baseline RAG pipelines, demonstrating improved navigation of hierarchical and multi-hop references.
CPGRec+ extends a balance-oriented gaming recommender by accounting for disparities in player-game interactions to prevent over-smoothing, balancing accuracy with diversity in personalized video game recommendations.
The paper addresses negative constraints in KGQA using schema-guided semantic matching and self-directed refinement to improve faithfulness and reduce hallucinations in answers requiring negative constraints.
MAGMA proposes a multi-graph agentic memory architecture for AI agents, representing memory items across orthogonal aspects such as semantic and temporal information to improve interpretability and alignment between query intent and retrieved evidence.
Entity State Tuning (EST) for temporal knowledge graph forecasting maintains persistent and evolving entity representations across time, addressing episodic amnesia and long-term dependencies to improve forecasting accuracy.