Explainable Machine Learning for Credit Card Customer Segmentation Using Clustering Algorithms

Qiu and Wang [1] found how machine learning can enhance the segmentation of credit card customers. There are many new procedures to do this Recency, Frequency, and Monetary (RFM) analysis and neural networks, among others that can increase the speed and accuracy of the segmentation. Using raw algorithms like Gaussian mixtures, DBSCAN, and K-means, a bank or any such financial institution can adapt the methods peculiarly to the spending patterns of its customers. Precise segmentation leads to economic stability and risk management through data-driven decision-making in the credit card business.

Talaat et al. [2] have provided ongoing literature research for describing the integration of Artificial Intelligence and Deep Learning for customer segmentation. Conventional approaches (such as RFM analysis) are incapable of capturing patterns that may not be easily recognized. Therefore, the DeepLimeSeg model integrates the neural network approach with the interpretability through the Lime-based technique. This will use the customers' behavioral data and demographics to segment the customer. DeepLimeSeg was tested on real-world datasets and demonstrated its efficiency in data-driven marketing decisions by outperforming traditional models on metrics such as Mean Squared Error (MSE), Mean Absolute Error (MAE), and R-square.

The use of k-means clustering for customer segmentation in RFM analysis was investigated by Gustriansyah et al. [3]. It is shown that k-Means is a useful tool for determining customer segmentation and optimizing marketing strategies. However, k-Means has drawbacks, such as sensitivity to the initial center of mass. Future avenues of research include investigating alternative clustering methods and improving k-Means to improve customer segmentation. This research contributes to the development of data-driven marketing strategies.

A comparison was made by Akshaya et al. [4] of the K-means and K-medoids clustering algorithms and reviews data mining techniques used to predict frequent item sets in credit card transactions. The study evaluates the performance of both algorithms using a dataset of 100,000 credit card transactions. The results show that K-medoids outperform K-means in terms of handling outliers and clustering accuracy. Future modifications to the clustering algorithms are proposed in this paper to enhance the prediction of frequent item sets in digital transactions.

Karo et al. [5] employed the K-means algorithm in the identification of different groups of credit card patrons concerning their usage habits. The following six clusters were identified: occasional customers, cash-using, luxury-good customers, timely payers, installment-payers, and recreational shoppers. Six was established as the most optimal number of clusters by the elbow method and the standard of the formed clusters was ascertained by the silhouette index. The work points out that the algorithm will only get better in the next time and it will be able to sort more variables.

Dawood et al. [6] examined the application of machine learning techniques (including K-means, improved K-means, fuzzy C-means, and neural networks) to the analysis of bank customer behavior. The dataset contains 30,000 customer records from a Taiwanese bank. After preprocessing, new labels are generated using an unsupervised clustering algorithm. Using these labels, the neural network achieved a maximum accuracy of 98.08%. With this approach, banks can reduce risk and increase profitability through improved customer analytics. Future research will explore deep learning techniques to further improve analytical accuracy.

The study by Umuhoza et al. [7] focuses on Egypt and performs behavior-based credit card segmentation of African users. It examines how credit card transaction analysis using unsupervised machine learning, more specifically k-means clustering, can classify users into four categories: ordinary people, fashion enthusiasts, executives, and limited consumers. The results could help financial institutions in Africa to better target their marketing efforts, satisfy more customers, and encourage the use of credit cards. The study also highlights the effectiveness of such segmentation in reducing marketing costs and increasing consumer loyalty.

The study by Jha [8] uses unsupervised machine learning to study credit card consumer behavior analysis. It uses clustering techniques such as K-means to group customers according to their spending habits based on a dataset of 9000 credit card holders. The analysis yields three groups: high-balance users, large consumers, and low-balance consumers. The study addresses data quality and ethical issues in the use of the model while highlighting how businesses can use data science to improve marketing strategies, detect fraud, and improve customer segmentation.

Zhou et al. [9] performed research and analysis of different machine learning techniques-based approaches for customer segmentation and personal credit rating estimation based on the dataset of 10,000 credit records. The study calls attention to the now swift shift towards digital payments and omnichannel economies, supported by data-driven approaches. LOF testing methods, gradient boosting, and BP neural networks that would help improve the accuracy of credit scoring will be discussed in this work. More precisely, this paper improved credit ratings, enhanced customer segmentation, and resolved issues of personal credit ratings in e-commerce by incorporating machine learning into conventional bank rating systems.

Rachman et al. [10] investigated Mini Batch K-means for credit card customers' clustering using a data set of 46,079 credit card transactions over three months. This study combined Recency, Frequency, and Monetary analysis with demographic, geographic, and behavioral data in an attempt to develop an improved customer segmentation. It proved that Mini Batch K-means outperformed other algorithms like K-means and BIRCH in runtime. Using the Elbow method, the optimal number of clusters proved to be four. This approach facilitates the study of target marketing by identifying customer loyalty and potential.

Abdulhafedh [11] applied K-means, Hierarchical Clustering, and Principal Component Analysis (PCA) to segment customers by using a dataset from 8,950 credit card holders. This shows that, according to the study with the various evaluation metrics, Silhouette and Dunn indexes, K-means outperforms Hierarchical Clustering. PCA reduces 17 features into 5 principal components, minimizing collinearity and enhancing the efficiency of clustering. Most importantly, PCA unveiled one more group of customers that increased segmentation accuracy. The study concludes that the combination of PCA with K-means improves the identification of distinct customer clusters and amplifies targeted marketing strategies.

Hung et al. [12] used HAC on 9,000 credit card holders featuring credit limits, purchases, and payments. The study preprocesses data by handling outliers and missing values, followed by the determination of three optimal clusters using techniques like the Elbow and Gap statistic. Results show three clear customer segments: low-credit, high-spending, and mixed patterns. It concludes that HAC is suitable only for a small dataset; it advocates cloud-based environments if one wants scalability. This enhances cluster analysis in the formation of marketing strategies in customer segmentation.

Yanık and Elmorsy [13] examined the segmentation of customers using SOMs and k-means clustering for data derived from credit card transactions, considering both consumption and demographic variables. By PCA, a dataset of 38,887 customers had their number of dimensions reduced for ease of clustering. The SOMs outperformed k-means, especially on the demographic feature representations, due to lower Davies-Bouldin indices. Among interesting features clustered: single, high-income people spend a lot on entertainment and insurance. It is recommended that future research apply these clusters to better forecasting and business strategies.

Figure 2 shows a full comparison of the clustering performance of different algorithms using the same 2D PCA projection, to provide a consistent view. Figure 2(a) shows the Distortion Score (Elbow Method). Here the “elbow” is at k = 4, which suggests four clusters may give the right amount of compactness without being overfit. Figure 2(b) shows the Calinski-Harabasz Score and it also has a peak at k = 4, suggesting the selection from the distortion score is reinforced. Based on this, the following commonly used clustering algorithms were applied to the dataset using k = 4 wherever appropriate.

Figure 2(c) illustrates the K-Means clustering outcomes where the clusters are notably separate in PCA space. It displays a successful clustering of four clear-cut clusters.

Figure 2(d) states the outputs from the DBSCAN clustering algorithm were similar to Figure 2(c). Including DBSCAN was important because it clustered multiple dense clusters and identified several points as noise (shown here with outliers), particularly at sparse locations in PCA space.

Figure 2(e) reflects the outputs from Hierarchical clustering with Ward linkage. It yielded similar clustering distributions to the K-Means clustering outputs, but the boundary definitions of the clusters as well as the internal compactness of the clusters differed slightly.

Figure 2(f) displays the outputs from Spectral Clustering. It captured the non-linear relationships between data better than the K-Means clustering outputs. The clusters are distinguishably discrete, but not entirely. There are slight overlaps which suggest that Spectral Clustering may be more parameter sensitive.

Table 1. Comparison of algorithms based on evaluation metrics

The data presented in the Table 1 indicates that the K-Means algorithm records the lowest Davies-Bouldin index among the three algorithms, suggesting that K-Means offers better clustering quality in comparison to the others. However, in terms of silhouette score, K-Means ranks second highest, indicating the presence of some overlapping clusters generated by this method. Additionally, the performance of the hierarchical clustering algorithm displays clustering quality results that are comparable to those of K-Means, although its Davies-Bouldin index is slightly elevated and its silhouette score is marginally lower. In contrast, DBSCAN demonstrates the poorest performance in terms of the Davies-Bouldin index, yet achieves the most favorable silhouette score among the algorithms assessed. The results from the Calinski-Harabasz index further illustrate that K-Means offers the highest value, signifying its superior performance and density relative to the other algorithms. In summary, it can be inferred that K-Means exhibits the highest clustering quality of the three algorithms, as evidenced by its lowest Davies-Bouldin index and slightly improved overlap of clusters when compared to hierarchical clustering.

Figure 3. Clustering results for K-means based on mean matrix

Clustering results for K-Means based on the mean matrix are shown in Figure 3 and SHAP summary plot provided in Figure 4 gives insight into which features are most impactful within the model's predictions—specifically, "BALANCE_FREQUENCY" and "BALANCE." For "BALANCE_FREQUENCY," higher values colored red mostly increased the model's prediction due to the positive SHAP value output, while the lower ones have a smaller impact, colored blue. For the feature "BALANCE," there is a more important interaction: high and low values may have either positive or negative contributions depending on the context of the prediction. The color bar further maps the color in between that shows the relationship in correspondence to the contribution coming from the value of that specific feature, ranging from the value in red, referring to higher values, through gray, to blue with its lower values.

Table 2. SHAP-based top feature contributions

Table 2 shows the top two features that influence each customer’s segment based on SHAP values. SHAP helps explain which features most impact a model’s prediction. Customer clusters are primarily influenced by BALANCE and BALANCE_FREQUENCY with BALANCE emerging as the most dominant feature across all groups. Figure 5 shows the predicted probabilities and feature contributions for Cluster 1 and Cluster 2 according to a clustering model. The prediction probabilities are identical between the two clusters, with Cluster 2 being the most likely assigned cluster at 0.59 and Cluster 1 being the second most likely at 0.22. Feature thresholds drive the decision-making in assigning Cluster 2: "BALANCE ≤ -1.49" contributes highly to the prediction, with a value of 0.21, whereas "-1.26 < BALANCE_FREQUENCY" adds contributions of smaller values, ranging between 0.07 and 0.08. These values are constant across the clusters for "BALANCE = -1.68" and "BALANCE_FREQUENCY = -1.08." Therefore, it seems from this that both features are relevant for cluster assignments, but "BALANCE" is driving the model more.