DC9 Project: Decentralised Graph Neural Networks: Adaptation, Training and Interpretability in Federated Environments

PhD Supervisor: Prof. Salvatore Cuomo (UNINA); Auxiliary supervisors: David Camacho (UPM), Dariusz Mrozek (SUT), Shen Yin (NTNU); R&D cooperation: ALMAWAVE

Objectives: Graph Neural Networks have emerged as powerful tools that can be used to process structured data, thereby capturing complex relationships and patterns. The primary objective of this research is to adapt and train GNNs within a federated learning framework. This will entail understanding how to partition the graph data across nodes, train GNNs without centralising this data and effectively aggregate the models or updates. Given the unique structure of graph data, special attention will be paid to the challenges in the data distribution and model aggregation that might not be present in traditional federated learning scenarios. Another significant objective will be to improve the interpretability of these decentralised GNNs. By investigating the decision-making processes of the GNNs that operate in a federated setting, the aim of the research is to make these models not only efficient but also transparent and understandable, which will bridge the gap between complex computations and actionable insights.

Expected Results: By the end of the project, a prototype of a GNN that is designed to operate in federated environments should be developed. This GNN will be adept at handling structured data across decentralised nodes, trained efficiently thus ensuring minimal information leakage. Alongside the technical model, a suite of tools or methodologies whose aim is to interpret the decisions and processes of the federated GNN is also anticipated. This would shed light on how the model processes the graph-structured data across the nodes and will offer explanations for its outputs. Furthermore, the research will provide insights into the challenges, both anticipated and unforeseen, of adapting GNNs to federated settings. It will also offer guidelines or best practices for other researchers who wish to embark on similar endeavours, thereby ensuring that the convergence of GNNs and federated learning is both efficient and insightful.

Applied research: The applied research in this project will develop and refine a Graph Neural Network (GNN) prototype suited for federated learning environments, focusing on handling decentralized graph-structured data. The project aims to innovate in training GNNs without centralizing data, ensuring efficient data processing with minimal information leakage. A significant focus will be on enhancing GNN interpretability in a decentralized context, making these models transparent and easily understandable. The project will culminate in a practical GNN framework, offering insights and best practices for GNN adaptation in federated settings, bridging theoretical research with practical applications in fields requiring complex data analysis.

Planned secondments: UPM(4 months); SUT(4 months); NTNU (4 months)                   

Enrolment in Doctoral degree: UNINA

Recruitment of Doctoral Candidate: https://euraxess.ec.europa.eu/jobs/290127