The data in banks is typically collected in data warehouses/data lakes and is organized in the form of relational databases (i.e. data tables together with links between them). For example, one database might store customer data, like name ,location, branch, age, etc., while a second table might store channel data such as channel, product, access time, and user, user being the connecting variable between the two tables, customers and channels.
To train machine learning on such connected data several approaches have been proposed, most recently an approach using graphs. In this approach relational databases are mapped into decorated graphs, with entries representing nodes, while relations label the edges in a graph.
With recent advancements in hardware and software, such graphs can be analyzed with deep learning, most prominently graph neural networks, a deep learning approach for analyzing large graphs and understanding their properties. Graph neural networks allow analysis of such relational databases (mapped to graphs), giving a completely novel, contextual understanding of data, coming from adding additional, relationally related data to predictive tasks.
Graph neural networks however have a critical disadvantage – they are very complex to execute. This is in part due to large size of graphs, as their complexity and complexity of their analytics grows exponentially with the number of nodes, and on other hand due to graphs being particularly incompatible with GPU processing due to their intertwined structure combining long-range and short-range interactions.
The shortcomings of graph neural networks can be addressed through quantum computing and quantum-inspired classical computing. Quantum computers batch data in a different manner compared to classical and greatly remedy GPU incompatibility, at the same time providing large computational power for exponentially complex problems of graph analytics with neural networks.
Quantum graph neural networks were introduced in 2019 as the quantum analogue of classical graph neural networks. They can be used for predictive analytics of graph data, finding missing features, graph evolution analytics, etc.. In banking the tasks can be translated to contextually segmenting customers, automated recommendations, predicting next purchase time, churn, etc., all now involving relational database data from the data storage, rather than a single table with limited data.
The goal of the project is to develop and deploy a quantum graph neural network for contextual customer segmentation on a relational database in financial institutions.
