Evaluating Deep Topology-Preserving Models for Behavioural Customer Segmentation in Open Banking Data : A Comparative Evaluation of Autoencoder and Self-Organizing Map Hybrid Architectures

Sammanfattning

The introduction of the European Union’s second Payment Services Directive (PSD2), which mandates secure, customer-consented access to bank account and transaction data, has enabled access to detailed, transaction-level financial data. This development creates new opportunities for behavioural customer segmentation based on observed financial activity rather than static demographic attributes. However, such data are high-dimensional, heterogeneous, and largely unlabeled, posing significant challenges for traditional clustering methods in terms of robustness, interpretability, and stability. This thesis evaluates the suitability of topology-preserving and deep representation learning approaches for behavioural customer segmentation in an open banking context. Using anonymised and categorised transactional data from nearly 10,000 individuals, four unsupervised segmentation pipelines are compared: direct constrained K-means clustering, a Self-Organizing Map (SOM)-based approach, a sequential autoencoder (AE) followed by SOM, and a topologyregularised AE–SOM architecture inspired by the Deep Embedded Self- Organizing Map (DESOM) framework. All pipelines are evaluated under identical conditions using internal clustering metrics, stability analysis, and qualitative interpretability through visualisation and cluster profiling. The results demonstrate a clear performance hierarchy, with the DESOMbased representation learning approach consistently achieving the most compact, well-separated, and stable clusters. While standalone SOMs provide strong visual interpretability in the original feature space, deep representation learning significantly improves structural cluster quality, albeit with some loss of feature-level transparency. Overall, the findings indicate that hybrid models combining autoencoders and Self-Organizing Maps are viable and effective methodologies for behavioural segmentation of PSD2-enabled transactional data, with trade-offs between interpretability and representational power that should be carefully considered in practical applications.