Showing 13 papers for 2026-05-04
HyperODE RCA presents a unified framework for root-cause localization in cloud-native microservice systems. It combines differentiable hypergraph attention to capture higher-order service interactions, latent ordinary differential equations to model irregular temporal dynamics, and multimodal cross-attention to fuse heterogeneous observability data, enabling fine-grained root-cause localization. The approach handles irregular observations and complex dependencies across multiple modalities.
We introduce the Causal Edge Classification Framework (CECF) to edge classification. Unlike previous work, CECF applies causal inference principles directly to edge features and their node-context, enabling modeling of how node features causally influence edge behavior in high-dimensional data. This leads to more accurate edge classification and insight into edge-level causality.
GD4 proposes Graph-based Discrete Denoising Diffusion for MIMO Detection. It tackles NP-hard MIMO detection, especially in under-determined systems with Nr < Nt. By performing graph-based discrete diffusion to progressively refine symbol estimates, GD4 achieves competitive performance with reduced inference iterations.
C-MTAD-GAT offers scalable context-aware graph attention for unsupervised anomaly detection in large-scale mobile networks. It processes thousands of heterogeneous KPI time series with a single shared model, achieving robustness to context shifts and nonstationarity through context-aware multivariate time-series analysis.
Soft Graph Diffusion Transformer (SGDiT) reframes MIMO detection as a noise-level-conditioned denoising process within a flow-matching framework. It progressively refines symbol estimates by diffusing from a Gaussian initialization toward the posterior conditioned on channel observations, using a graph-based diffusion transformer with adaptive layers.
Graph Rewiring in GNNs to Mitigate Over-Squashing and Over-Smoothing: A Survey. GNNs often suffer from over-squashing and over-smoothing due to topology and message passing. This survey reviews graph rewiring techniques that modify graph structure to improve information flow, categorizing methods and providing guidelines for when to apply them to mitigate these issues.
Adaptive Node Feature Selection For Graph Neural Networks. We propose an adaptive node feature selection approach that identifies and removes redundant features during GNN training. By measuring feature contribution to the model output in a data-, model-, and task-agnostic way, it reduces dimensionality and enhances interpretability without requiring task-specific heuristics.
When Structure Doesn't Help: LLMs Do Not Read Text-Attributed Graphs as Effectively as We Expected. This study finds that large language models do not read or reason over text-attributed graphs as effectively as plain text or unstructured prompts when using graph structure. Experiments show limited gains from graph templates or GNN-encoded structure, highlighting the need for better integration methods or prompting strategies to harness graph information.
Semantic Level of Detail for Knowledge Graphs: Discovering Abstraction Boundaries via Spectral Heat Diffusion. We introduce Semantic Level of Detail (SLoD) for knowledge graphs to continuously control abstraction levels. By applying spectral heat diffusion, SLoD discovers gradual abstraction boundaries and enables a continuous zoom across knowledge graph granularity with theoretical guarantees. This supports principled navigation and querying across multiple detail levels.
Empowering Heterogeneous Graph Foundation Models via Decoupled Relation Alignment. We propose Decoupled Relation Alignment to empower graph foundation models on multi-domain heterogeneous graphs. By decoupling alignment across relation types, the method avoids type collapse and relation confusion, preserving type-specific semantics while aligning cross-domain knowledge. This improves cross-domain generalization and topological fidelity in heterogeneous graphs.
Controllable Logical Hypothesis Generation for Abductive Reasoning in Knowledge Graphs. We address controllable hypothesis generation for abductive reasoning on knowledge graphs, reducing redundancy and irrelevance in plausible hypotheses. The approach introduces mechanisms to constrain and steer hypothesis generation, improving practical utility for domains like clinical diagnosis and scientific discovery.
G-reasoner: Foundation Models for Unified Reasoning over Graph-structured Knowledge. We present G-reasoner, an approach that equips foundation models with unified reasoning capabilities over graph-structured knowledge. Addressing limitations of static parametric knowledge, it leverages graph-augmented retrieval and reasoning (GraphRAG) to better integrate structured graph information into LLMs.
Robust Multimodal Recommendation via Graph Retrieval-Enhanced Modality Completion. We study robust multimodal recommendation when modalities may be missing and propose modality completion via graph retrieval-enhanced methods to reconstruct missing features and produce modality-complete graphs for downstream tasks. This improves recommendation performance under modality incompleteness.