Foundation models have transformed language, vision, and multimodal AI, but equivalent progress for graph-structured data remains at an early stage. GFMs [1, 2] aim to learn reusable representations across graph domains and tasks, but major challenges remain in transferability, graph heterogeneity, structural alignment, scalability, and evaluation. Recent surveys describe the fi eld as promising but still immature , with open questions around backbone architectures, pretraining strategies, adaptation mechanisms, and generalisation across diverse graphs [3, 4]. This gap is highly relevant to fi nancial services, where many important problems are relational rather than purely textual or tabular. Clients, accounts, legal entities, counterparties, transactions, documents, policies, controls, etc. form complex networks. Current enterprise AI approaches often rely on standard retrieval-augmented generation, tabular analytics, or task-specifi c graph models. These methods may struggle to reason over multi-hop relationships, provide reliable evidence, or generalise across diff erent business domains. GraphRAG has already been explored for fi nancial fraud detection [5] because fraud evidence often lies in relationships across fragmented data rather than individual records alone. However, the broader question of whether GFMs can become reusable AI infrastructure for fi nancial services remains insuff iciently examined.
The project will develop a conceptual architecture showing how GFMs could interact with enterprise knowledge graphs, LLM interfaces, GraphRAG-style retrieval, explanations, and governance controls. This will be compared with conventional RAG, task-specific graph models, and standard knowledge graph querying to identify where GFMs provide genuine added value.
This project is NOT suitable for undergradaute students, unless you have have expertise in graph machine learning, knowledge graphs, and large language model systems, with the ability to critically assess emerging methods such as Graph Foundation Models, GraphRAG, graph transformers, and graph instruction tuning.
[1]: DOI: 10.52202/085713-1222
[2]: DOI: 10.52202/079017-3412
[3]: DOI: 10.1109/TPAMI.2025.3548729
[4]: DOI: 10.48550/arXiv.2402.02216
[5]: DOI: 10.1145/3786484.3786522