Fan Pan

In cloud-network convergence environments, Virtual Network Embedding (VNE) confronts three fundamental challenges: resources heterogeneity, services dynamicity and policies diversity. Traditional exact algorithms, (meta)heuristic methods and learning-based approaches are respectively con strained by computational complexity, limited generalization capability and data dependency, rendering them inadequate for addressing these challenges effectively. Although Large Language Models (LLMs) demonstrate remarkable capabilities in semantic understanding and reasoning, their direct application to VNE still suffers from difficulties in cross-modal coordination, insufficient solution reliability and the inability to quantify business intents. To address these limitations, we propose FLEXVNE, an LLM-driven flexible VNE framework that incorporates three key innovations: (i) a coordinated feature encoder that fuses network graph structure with natural language descriptions to precisely characterize heterogeneous resources; (ii) a VNE head equipped with dynamic masking and hierarchical pol icy networks, embedding hard constraints directly into the decision process to ensure feasible and reliable embedding solutions; (iii) an intent-driven policy controller that transforms ambiguous natural language requirements into optimizable objective weights, guiding policy adaptation to diverse business preference. We conduct comprehensive evaluations of FLEXVNE on a simulation platform. Experimental results demonstrate that across topologies of varying scales and constraint intensities, FLEXVNE significantly outperforms all baseline schemes across all evaluation metrics. Furthermore, FLEXVNE successfully accommodates various user-defined intents, offering a promising paradigm for building intelligent and adaptive cloud-network resource orchestration systems.

Cloud-network convergence represents a pivotal architectural evolution in modern communication networks, enabling unified orchestration of computing, storage, and network resources. Within such environment, trust evaluation plays an essential role in ensuring secure resource scheduling and reliable service deployment. While graph representation learning (GRL) has emerged as a powerful paradigm for trust evaluation by capturing complex relational patterns among entities, existing approaches still face three fundamental challenges: resource heterogeneity, data sparsity, and insufficient interpretability. To address these challenges, we propose LAGTrust, a Language-Augmented Graph representation learning framework for trust evaluation in cloud-network convergence. LAGTrust integrates large language models (LLMs) with graph neural networks through three synergistic components: (i) a Heterogeneous Feature Representation module that leverages LLMs to project multi-dimensional resource attributes into a unified semantic space while employing spatio-temporal graph neural networks to capture structural and behavioral dynamics; (ii) an Adaptive Feature Enhancement module that performs hierarchical LLM-driven augmentation for semantic completion, relationship inference, and profile enrichment; and (iii) a Semantic Explanation Generation module that synthesizes multi-source evidence through causal reasoning to produce interpretable trust assessments. Extensive experiments on real-world and synthetic topologies demonstrate that LAGTrust consistently outperforms baseline methods across multiple evaluation metrics, particularly under cold-start scenarios, validating its superior accuracy, robustness, and interpretability for trust evaluation in cloud-network convergence environments.