Businesses are aware of the importance of customer relationship management and strive to gain insights into customers and their needs. Clustering and prediction are widely used data mining methods for understanding customer behavior. In this paper, we introduce an intelligent approach that utilizes time series clustering to analyze customer behavior. To segment customers, we first construct time series of their behavior and then perform preprocessing on the data. Subsequently, a set of informative features reflecting the characteristics of each time series is extracted. These features are ranked using the Laplacian Score method and employed in clustering. The proposed method is applied to time-stamped transaction data of bank customers. After constructing a customer behavior series and extracting features, four informative features —variance of all points in the time series, entropy, spikiness, and lumpiness —are selected for clustering. Customer clustering is performed using four state-of-the-art clustering algorithms: k-medoids, k-means, Fuzzy C-Means clustering (FCM), and Self-Organizing Map (SOM) algorithms. The results demonstrate that, among various clustering methods, the k-medoids algorithm outperforms others. It divides customers into four clusters with a Silhouette metric of 0.6378.
