Efficient HMM-Based Interpretable Customer Representation for Customer Segmentation in Banking

Abstract Customer segmentation is a fundamental component of banking decision-support systems, supporting applications such as targeted marketing, risk management, fraud detection, and strategic planning. Existing approaches generally combine feature extraction with clustering, relying on either static or dynamic \textit{Recency--Frequency--Monetary} (RFM) representations or deep learning models. However, RFM-based methods often fail to capture the sequential dynamics of customer behavior, while deep learning approaches typically produce latent representations that are difficult to interpret and frequently require additional dimensionality reduction for visualization. To address these limitations, this paper introduces the \textit{Sequential Amounts and Sequential Intensity} (SASI) framework, an interpretable customer representation method based on Hidden Markov Models (HMMs). Unlike conventional deterministic approaches, HMMs provide a probabilistic framework for modeling sequential variations in transaction amounts and transactional intensity. The meta-data extracted from the learned HMMs are transformed into a two-dimensional embedding, whose axes directly reflect these two behavioral dimensions. Experiments conducted on $1{,}503$ and $2{,}039$ customers from the synthetic \textit{Nigerian Retail Banking Transactions (Nigerian)} and the real-world \textit{PKDD'99 (Berka)} datasets demonstrate the effectiveness of SASI using the \textit{Silhouette Coefficient} (SC), \textit{Davies--Bouldin Index} (DBI), and \textit{Calinski--Harabasz Index} (CHI). When SASI is compared with the competing methods, it achieves the second-best $\mathrm{SC}$ ($0.650635$), the best $\mathrm{DBI}$ ($0.418718$), and the best $\mathrm{CHI}$ ($6578.11$) on the \textit{Nigerian} dataset. On the \textit{Berka} dataset, it attains the best $\mathrm{SC}$ ($0.507139$), the best $\mathrm{DBI}$ ($0.500209$), and the second-best $\mathrm{CHI}$ ($3309.01$). These results establish SASI as an effective operational framework, combining competitive clustering performance with fully interpretable and directly visualizable customer representations.