A new sorting-based classification approach to digital customer segmentation

Abstract

As digitalization continues to expand, firms seek to adapt in order to reach customers and respond to their needs. The telecommunications sector, as one of the industries with the largest customer bases, has similarly focused on developing strategies to remain competitive in this environment. This thesis aims to define the concept of digitalization using telecommunications data and to provide sector-specific recommendations based on this definition. In the first stage, academic studies on digitalization, digital customers, and digital channels were reviewed across the telecommunications sector as well as other industries. The literature shows that research on digitalization is limited and that existing definitions are generally narrow, most often relying on transaction volumes and data usage. Studies from banking and e-commerce were also considered, including segmentations based on demographic characteristics or channel preferences. Based on these findings, this study adopts a broader view of digitalization, incorporating not only transaction counts or channel choice but also customers' spread across digital channels, behavioral diversity, and depth of digital activity. Accordingly, digital classes were constructed and analyzed in detail. Following the establishment of the definition of digitalization and the research objectives, the methodological framework was defined and the implementation phase was initiated. A sample of fifty thousand customers was selected from real telecommunications data, and a total of forty-five variables were identified based on those frequently used in the literature and in telecommunications practice. These variables were selected to capture customers' total data usage, usage patterns across different time periods, generated revenue, physical mobility, levels of social media usage, and the frequency of visits to websites and applications across various domains such as gaming, public services, and news. After the variables were selected, numerical data were transformed into categorical form using the Fisher–Jenks Natural Breaks method in order to enable their use in the classification methodology. This method identifies breakpoints that maximize between-class variance while minimizing within-class variance, and the optimal number of categories was determined using the Goodness of Variance Fit (GVF) metric. Following the categorization, the relationship between the resulting categories and the existing digital segmentation structure was examined using chi-square and Cramér's V tests to ensure that the selected variables were meaningful and consistent with the current framework. Subsequently, the ELECTRE TRI-B multi-criteria ordered classification method was applied. Prior to classification, variable weights were calculated using the entropy method, which assigns higher weights to variables with greater variability. Navigation website visits and text- and video-based social media data usage emerged as the most discriminative variables. Customers were then classified into six classes, consistent with the number of segments in the existing digital segmentation structure, enabling a direct comparison between the two approaches. Subscribers were assigned to digital classes on a monthly basis over a six-month period. Segment distributions were stable across periods, with no significant seasonality effects. Approximately 36% of customers were concentrated in the third class, followed by the fourth class at around 32%, while the first class represented the smallest segment at roughly 2%. Profiling analysis revealed clear differences across classes. The first class exhibited almost no data usage, while the second class generated the lowest revenue and had the lowest package prices. In higher classes, increases in digitalization were accompanied by higher revenues, fees, and usage volumes. The second class consisted of the oldest customers and had the longest subscription tenure, whereas higher segments were characterized by younger customer profiles. Despite minimal usage, the first class generated higher revenue and had a younger customer base compared to the second class. Finally, inter-period transition analysis was conducted to examine mobility between segments. Transition matrices showed that mobility was more pronounced in upper segments, with most subscribers moving to adjacent classes between periods. The high level of movement observed in high-revenue segments represents a critical insight for segment management and strategic decision-making. After analyzing intra-segment dynamics, comparisons were conducted between the existing segmentation structure and the ELECTRE TRI-B-based classes to identify the distinguishing features of the new structure. Initially, similarity levels between the six classes in both classification systems were examined. No significant seasonality effects were detected in similarity analyses conducted for each period. While the fourth, fifth, and sixth classes in both structures exhibited high similarity across all combinations, the lowest segments of the two classification systems showed low similarity. The second and third classes also demonstrated high similarity with each other. These findings constitute one of the most important indicators that the ELECTRE TRI-B-based classes differ substantially from the existing segmentation structure. The classes were also compared in terms of income levels. In the existing segmentation, average revenue per subscriber increased progressively across higher segments. In contrast, within the ELECTRE TRI-B classes, the lowest average income was observed in the first class, followed by the second and third classes. Moreover, income differences among upper segments were considerably more pronounced in the ELECTRE TRI-B classification. Age was another variable used for comparison, and notable differences were observed in this dimension as well. In the existing segmentation structure, the average ages of the first three segments were relatively similar, and customers in higher segments tended to be younger. Conversely, in the ELECTRE TRI-B classes, the second class exhibited a distinctly higher average age, while the average ages of upper segments were slightly lower than those of the corresponding upper segments in the existing structure. Another comparison was conducted based on total data usage. While a steadily increasing pattern was observed across segments in the existing classification, in the ELECTRE TRI-B structure the data usage of the second class was only marginally higher than that of the first class, whereas the upper classes displayed significantly higher average usage levels. In parallel with total data usage, similar trends were observed in social media platform usage. In the final stage, the two classification structures were compared in terms of their ability to explain income and data usage using regression analysis. In this context, segments were included in the models as independent variables, and revenue per subscriber was selected as the dependent variable. Although the explanatory power of both models was relatively low, the R² values indicated that the ELECTRE TRI-B classes exhibited higher explanatory power for income compared to the existing segmentation structure. It should be noted that revenue is influenced by numerous external factors, including promotional campaigns, price changes, competitors' offers, and brand perception. The same analytical steps were applied to total data usage, and similarly, the ELECTRE TRI-B classes demonstrated greater explanatory power than the existing classification. These findings indicate that the proposed segmentation approach provides a stronger framework for explaining both revenue and usage behaviors compared to the current structure. In conclusion, the method applied in this study resulted in the development of a digital segmentation structure that clearly differs from existing classification approaches. By combining real telecommunications customer data with variables selected from multiple dimensions, statistical methods, and expert judgment, the proposed segmentation offers a meaningful contribution to both industry applications and the academic literature. In addition to the segmentation itself, detailed profiling of the newly formed classes and the analysis of inter-period transitions provide valuable insights that can be used as inputs for segment management and strategic decision-making processes. While the study was conducted using data from existing postpaid subscribers, extending the analysis to newly acquired and prepaid customers would increase the generalizability of the proposed structure. Moreover, testing the results through real business actions would be essential to assess the practical value of the segmentation and to identify potential areas for improvement. Future research may further deepen the analysis by experimenting with different numbers of classes. In addition, although churn behavior could not be examined due to data limitations, incorporating such analyses in future studies would offer important insights into segment sustainability and customer lifecycle dynamics.