Comparative Unsupervised Clustering Approaches for Customer Segmentation

Abstract

Machine learning-driven studies to get potent insights about customers are essential for the business world to grow as they achieve smarter in marketing and sales activities. Finding the consociate patterns of customer interaction activities leads to finding sensible segments. By this, strategists can reach out to different groups of customers with customized services, offers and plans. However, although clustering algorithms are reliable by virtue of them being competent studies, not all of them fit the studied domain. In this study, six well-known clustering algorithms with different parameters are applied to real-life customer purchase history data. The outcomes are compared, and the density distribution of data features in created clusters are visualized. Thus, it is possible to see the role of each selected feature on the differentiation of clusters. The cluster labels of data points (customers) are mapped in pairs of algorithms. As a result, the similarities and differences in clusters created by different algorithms are more straightforward to catch. Moreover, in addition to labeling data points with class labels, a hybrid approach is presented to obtain information about class label probabilities by fitting the support vector classification model. The proposed study gives promising results in understanding how different clustering algorithms fit the customer data and stands out with multi-sides evaluation and comparison experiments.