💻 computer science

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

This paper proposes the SASI framework, an interpretable Hidden Markov Model-based approach that effectively captures sequential customer behavior dynamics for banking segmentation, achieving competitive clustering performance and direct visualizability on both synthetic and real-world datasets.

Original authors: KELY MAXIME MOTUE DJOKO, SYLVAIN ILOGA

Published 2026-11-11

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: KELY MAXIME MOTUE DJOKO, SYLVAIN ILOGA

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the vast, humming ecosystem of modern banking, every swipe of a card and every digital transfer leaves a trace. For decades, financial institutions have tried to make sense of these traces by grouping customers into categories, a practice known as segmentation. This sorting is not merely an administrative task; it is the foundation for how banks decide who gets a loan, who receives a special offer, and who might be at risk of fraud. Traditionally, this sorting relied on a simple, static snapshot of a person's activity: how recently they spent money, how often they did so, and how much they spent in total. While easy to understand, this method treats a customer's history like a single photograph, ignoring the rhythm and flow of their daily financial life. It cannot distinguish between a person who spends steadily over a year and someone who makes one massive purchase and then disappears for months, even if their total spending is identical. As banking becomes more digital and complex, the need to understand the actual sequence of events—the story behind the numbers—has become urgent.

A team of researchers from universities in Cameroon and France has proposed a new way to tell that story. They developed a system called SASI, which stands for Sequential Amounts and Sequential Intensity. Instead of taking a static snapshot, this approach listens to the rhythm of a customer's transactions over time. The researchers realized that to truly understand a bank customer, one must look at two specific things: the size of the money moving in and out, and the time gaps between those movements. To capture this, they used a mathematical tool called a Hidden Markov Model. In simple terms, this tool acts like a probabilistic map that learns the hidden patterns in a sequence of events. It does not just count transactions; it learns the likelihood of a customer making a large deposit after a long wait, or a small withdrawal following a quick series of purchases. By analyzing these patterns, the system translates a complex history of hundreds of transactions into a single, easy-to-read point on a two-dimensional graph.

The researchers tested this idea on real-world data from two different sources. One was a synthetic dataset mimicking the banking habits of 1,503 customers in Nigeria, capturing the specific nuances of mobile banking and regional spending patterns there. The other was a historic, real-world dataset from a Czech bank, containing six years of transaction records for 2,039 customers. In both cases, the SASI system successfully grouped customers into distinct clusters based on their behavioral rhythms. The results were striking. On the Nigerian data, the system created groups that were highly distinct from one another, with a level of internal consistency that outperformed many existing methods. On the Czech data, it achieved the highest possible score for grouping accuracy among the methods tested. Crucially, unlike many modern artificial intelligence systems that produce results hidden inside complex, unreadable code, the SASI system produced a clear, visual map. A bank analyst could look at the graph and immediately understand what each group represented: perhaps a "very intensive contributor" who moves money frequently and in large amounts, or a "dormant consumer" who rarely interacts with the bank.

The study also addressed a common problem in data science: the trade-off between accuracy and understanding. Many advanced computer models can group customers very well, but they do so by compressing information into high-dimensional spaces that humans cannot visualize without further processing. These models often act as "black boxes," where the reason for a specific grouping is unclear. The SASI framework avoids this by design. It forces the data into a two-dimensional space where the horizontal axis represents the sequential amounts of money and the vertical axis represents the intensity or frequency of the activity. This means the final result is not just a list of numbers, but a picture that tells a story. The researchers found that this approach was not only accurate but also efficient. Processing the thousands of customers took less than an hour on a standard computer, and when the work was split across multiple processors, the time dropped significantly, suggesting the method could scale to handle millions of customers in a real banking environment.

However, the researchers are careful to note the boundaries of their work. The system currently relies on the assumption that a customer's behavior over a fixed period is a reliable indicator of their future actions. It also simplifies the data by normalizing transaction amounts, which helps in comparing different customers but might be sensitive to extreme outliers. Furthermore, the current version groups customers into spherical shapes, which works well for many cases but might miss more complex, irregular patterns in customer behavior. Despite these limitations, the findings suggest a powerful shift in how banks can view their clients. By moving from a static count of transactions to a dynamic understanding of behavioral sequences, financial institutions can gain a clearer, more human-readable picture of their customers. This clarity allows for better decision-making, whether it is tailoring a financial product to a specific lifestyle or spotting unusual activity that signals fraud. The work demonstrates that it is possible to build sophisticated analytical tools that remain transparent and interpretable, bridging the gap between complex data and human understanding.

Technical Summary: Efficient HMM-Based Interpretable Customer Representation for Customer Segmentation in Banking

1. Problem Statement

Customer segmentation is a cornerstone of banking decision-support systems, facilitating targeted marketing, risk management, and fraud detection. However, existing approaches suffer from significant limitations regarding the representation of customer behavior:

Static RFM (Recency–Frequency–Monetary): While interpretable and efficient, static RFM aggregates transactions into single statistics, ignoring the sequential organization and temporal dynamics of customer activity. Customers with identical RFM triplets may exhibit vastly different behavioral patterns (e.g., early vs. late activity bursts) that RFM fails to distinguish.
Dynamic RFM and Deep Learning: Dynamic RFM captures temporal evolution but still relies on aggregated statistics within sliding windows. Deep learning models (e.g., VAEs, LSTMs, Transformers) capture complex sequential dependencies but often produce high-dimensional latent representations that are difficult to interpret, require additional dimensionality reduction for visualization, and demand substantial computational resources.

There remains a gap for a method that simultaneously offers interpretability, low-dimensional visualization, compatibility with conventional clustering, and computational efficiency while capturing the sequential dynamics of both transaction amounts and transaction intensity.

2. Methodology: The SASI Framework

The authors propose the Sequential Amounts and Sequential Intensity (SASI) framework, which utilizes Hidden Markov Models (HMMs) to generate an interpretable, two-dimensional customer embedding.

2.1 Data Preprocessing and Markov Chain Construction

For a customer $U$ with a chronological transaction window $w = (d_1, a_1), \dots, (d_T, a_T)$ , the framework performs the following transformations:

Window Extension: Two dummy zero-amount transactions are inserted at the start ( $d_{start}$ ) and end ( $d_{end}$ ) of the observation period to account for inactivity periods.
Pseudo-Markov Chain ( $\hat{\delta}$ ): The window is initially mapped to a pseudo-Markov chain where states are dates and observations are transaction amounts.
Discretization and Normalization:
- Intensity (States): Dates are converted into inter-transaction delays ( $D(d_i, d_j)$ ) and discretized into 7 activity levels (from "same day" to "beyond 6 months"). These levels form the hidden states ( $N=7$ ) of the HMM.
- Amounts (Observations): Transaction amounts are normalized by the customer's maximum absolute transaction amount and discretized into $2M+1$ intervals (positive for deposits, negative for withdrawals, zero for null). These form the observation symbols.

2.2 HMM Initialization and Training

Unlike random initialization, SASI uses a data-driven initialization strategy:

The transition matrix ( $A$ ), emission matrix ( $B$ ), and initial state distribution ( $\pi$ ) are computed directly from the empirical frequencies observed in the discretized Markov chain.
A smoothing constant ( $\epsilon$ ) is applied to prevent zero probabilities.
The Baum–Welch algorithm (Expectation-Maximization) is then used to train the HMM ( $\lambda$ ) to probabilistically learn the sequential patterns of the customer's behavior.

2.3 Meta-Data Extraction and Scoring

The long-term behavioral characteristics are summarized using two stationary distributions derived from the trained HMM:

Stationary State Distribution ( $\phi$ ): Represents the proportion of time spent in each activity intensity state.
Stationary Symbol Distribution ( $\bar{\phi}$ ): Represents the proportion of time spent observing specific normalized transaction amounts.

These distributions are transformed into two interpretable scores:

Score $x$ (Sequential Amounts): A weighted sum of $\bar{\phi}$ , distinguishing between deposits and withdrawals with quadratic separation to emphasize magnitude differences. The result is scaled by the maximum transaction amount and mapped to a logarithmic scale for scale invariance.
Score $y$ (Sequential Intensity): A weighted sum of $\phi$ , emphasizing the frequency of activity states. The result is scaled by the total number of transactions and normalized to the interval $[0, 100]$ .

The final output is a 2D embedding $(x, y)$ for each customer, which can be directly visualized and clustered (e.g., using K-means) without further dimensionality reduction.

3. Key Contributions

The paper outlines four primary contributions:

Interpretable HMM-Based Representation: A novel method that jointly captures sequential transaction amounts and intensity using only transaction histories, avoiding hand-crafted aggregate statistics.
Two-Dimensional Behavioral Embedding: A direct 2D representation that preserves semantic interpretability (where axes correspond to amount and intensity) and allows for immediate visualization.
Pipeline Compatibility: The representation is fully compatible with conventional clustering algorithms, enabling seamless integration into existing segmentation pipelines.
Empirical Validation: Demonstration of the framework's effectiveness and computational efficiency on two open datasets (synthetic Nigerian and real-world Berka).

4. Experimental Results

The framework was evaluated on:

Nigerian Dataset: 1,503 customers (synthetic, 2 years of data).
Berka Dataset: 2,039 customers (real-world Czech bank, 6 years of data).

4.1 Clustering Performance

Using K-means with $K=5$ clusters, SASI was compared against Static RFM, Dynamic RFM (with various aggregations and LSTM encoders), and VAE-based representations.

Nigerian Dataset: SASI achieved the best DBI (0.418718) and best CHI (6578.11), and the second-best Silhouette Coefficient (0.650635).
Berka Dataset: SASI achieved the best SC (0.507139), best DBI (0.500209), and second-best CHI (3309.01).

Note: While some deep learning models (e.g., LSTM2, VAE2) showed higher SC or CHI scores, the authors note these results were misleading due to the generation of duplicate embeddings (degeneracy) or highly unbalanced clusters, rendering them less reliable for practical segmentation.

4.2 Interpretability and Visualization

The resulting 2D embeddings revealed a clear hierarchical structure:

The x-axis separated customers into "consumers" (negative values) and "contributors" (positive values).
The y-axis further subdivided these groups based on activity intensity (e.g., very intensive, regular, dormant).
This allowed for the creation of semantically meaningful profiles (e.g., "very intensive consumer") directly from the plot.

4.3 Computational Efficiency

Sequential Execution: Processing 1,503 Nigerian customers took ~1,075 seconds; 2,039 Berka customers took ~597 seconds (approx. <1 second per customer).
Parallel Execution: Using an MPI-based parallel implementation on 8 processors reduced execution time by a factor of ~3.1, demonstrating scalability for large datasets.

5. Significance and Claims

The paper positions SASI as an effective operational framework that bridges the gap between the simplicity of RFM and the expressiveness of deep learning. Its significance lies in:

Balancing Performance and Interpretability: Unlike deep learning models that produce "black box" latent spaces, SASI provides a fully interpretable 2D space where axes have direct semantic meaning (amount and intensity).
Capturing Sequential Dynamics: By modeling the order and timing of transactions via HMMs, SASI distinguishes customers that static RFM would treat as identical.
Operational Viability: The framework is computationally efficient, scalable via parallelization, and does not require complex hyperparameter tuning or dimensionality reduction steps, making it suitable for integration into banking decision-support systems.

The authors acknowledge limitations, such as sensitivity to extreme transaction amounts during normalization and the current reliance on K-means (which assumes spherical clusters), but assert that the framework provides a robust, transparent, and efficient solution for customer segmentation in banking.