PhD Position – Cognitive Network Observability: Bridging Graph Neural Networks and Large Language Models for Generalizable and Explainable Network Tomography

Type de contrat : CDD Contexte et atouts du poste The PhD will be hosted at the IRISA/Inria Centre at the University of Rennes, a major and recognized player in the field of digital sciences. The centre comprises more than thirty research teams and is at the heart of a rich R&D and innovation ecosystem. The position will be based on the Beaulieu Scientific Campus of the University of Rennes, a medium-sized town with an intense student life (approximately 25% of the population). Rennes is a dynamic, lively city and a major centre for higher education and research in France. The PhD will be hosted by the ERMINE project-team (Measuring and Managing Network operation and economics), a joint team between Inria and IRISA (Institut de Recherche en Informatique et Systèmes Aléatoires), in collaboration with the ADOPNET team. Working conditions : - Partial reimbursement of public transport costs - 7 weeks of annual leave plus 10 extra days off (RTT – statutory reduction in working hours) - Possibility of teleworking (90 days per year) and flexible organization of working hours - Access to training, cultural and sports activities, and vocational training The GENIE ANR project (Network Optimization and Generative Intelligence Ecosystem) aims to revolutionize network infrastructure management by combining the strengths of Large Language Models (LLMs) with network domain-specific expertise. The project, funded by the French National Research Agency (ANR), involves partners including University of Rennes, IMT, LEAT, L3i, and LabHC. GENIE addresses the limitations of conventional techniques by enabling an interpretable and adaptable approach to network optimization and automated management. The project will design an LLM pipeline for network management, studying collaborative and scalable LLM-based strategies that enable parallel processing, including a consensus mechanism to maintain effective decision-making. Mission confiée Context and Problem Statement As networks evolve to support ultra-reliable, low-latency communications (URLLC) and complex virtualization paradigms (e.g., 5G/6G Network Slicing), they are becoming increasingly dynamic and opaque. This creates a severe "visibility gap" where operators struggle to diagnose faults without direct administrative access to the underlying infrastructure. Network tomography provides a vital tool to achieve observability by inferring hidden link metrics from end-to-end measurements. However, traditional algebraic and statistical methods suffer from rigidity and scalability issues. Recent advancements have demonstrated that Machine Learning, specifically Relational Graph Convolutional Networks (RGCNs) operating on line graphs, can learn shared link relations and generalize monitor selection. Despite these successes, significant open challenges remain. First, in contrast to most existing literature that relies heavily on synthetic or idealized simulations, there is a critical need to evaluate and generalize these models across entirely different, unseen topologies using real or highly realistic network data. Furthermore, handling non-additive metrics (such as congestion or binary link failures) under these realistic conditions requires more sophisticated architectures. Second, while Graph Neural Networks (GNNs) can accurately infer *where* a degradation occurs mathematically, they lack the administrative and operational context to explain *why* it is happening. This thesis proposes a novel framework that bridges the mathematical inference capabilities of GNNs with the contextual reasoning of Large Language Models (LLMs) to achieve true cognitive network observability. It is conducted within the framework of the GENIE ANR project. Assignments: The PhD student will be responsible for conducting full-time research activities centred on the theme of the thesis: cognitive network observability through the integration of Graph Neural Networks and Large Language Models. The specific assignments include: 1. Advanced GNN Development: Design and implement Relational Graph Convolutional Networks (RGCNs) and novel message-passing paradigms capable of inferring non-additive network metrics and achieving zero-shot transfer to unseen topologies. 2. Graph-RAG Framework Design: Develop a Retrieval-Augmented Generation (RAG) framework tailored for network graphs, including the construction of a vector database containing historical incident tickets, BGP routing logs, maintenance schedules, and vendor documentation. 3. LLM Integration and Evaluation: Integrate the GNN inference pipeline with an LLM equipped with the Graph-RAG architecture, and evaluate the system's ability to generate human-readable, context-aware root cause diagnostics. 4. Experimental Validation: Conduct extensive experiments on real or highly realistic network data to validate the generalizability and robustness of the proposed framework across diverse, unseen topologies. 5. Scientific Dissemination: Write and publish res