Segmentação de Usuários Usando K-means e Usabilidade de Aplicativos de Vendas

Ana Carolina Sabbadin; Juliano Schimiguel

DOI: 10.22167/2675-6528-202602650

Article derived from a Course Conclusion Work (TCC), with content based on the student’s original work and adapted to the editorial format of the E&S Magazine with the support of the ResumeAI tool, an artificial intelligence solution developed by Instituto Pecege for textual synthesis and organization.

Summary

Usage patterns of a sales routine support application in a corporate context were analyzed, considering the growing relevance of mobile tools for sales teams and the impact of usability on productivity and technological adoption. The objective was to segment users based on engagement and usability metrics extracted from system usage logs, using the K-means clustering algorithm. The methodology employed a quantitative and applied case study, with anonymized data from 21,878 salespeople from four countries (Brazil, Mexico, Colombia, and Argentina), in the period from January to December 2025. Steps included preparation, MinMax normalization, K-means application, and validation by Elbow and Silhouette. The results revealed three distinct profiles – occasional, regular, and advanced – differentiated mainly by the intensity, depth, and diversity of interactions, rather than by access regularity, which reinforced the distinction between frequency and quality of use. The contributions provided subsidies for personalization of engagement strategies, design optimization, and data-driven management in large-scale commercial environments.

Keywords: Unsupervised learning; Clustering; User engagement; Sales force support systems; Usability.

1. Introduction

The growing ubiquity of mobile applications has transformed the way individuals and organizations perform tasks, offering unprecedented convenience and accessibility. Users often prefer mobile applications over traditional web interfaces due to greater agility and ease in performing activities (Olanrewaju et al., 2020). This trend is corroborated by Gartner (2021), which highlights the rise of mobile applications as essential tools for both consumers and employees, enabling access to services and information anytime and anywhere.

In this scenario, large companies have been investing significantly in the development of applications, not only for their clients but also for their internal teams, such as the sales force. Organizations increasingly rely on mobile applications to support internal operations and engage customers, especially in sales and service departments (Ghezzi et al., 2018). Such mobile platforms for employees promote greater coordination and efficiency in processes (Bharadwaj et al., 2013), facilitating task execution and providing real-time guidance, which results in increased productivity and better decision-making (Venkatesh et al., 2012; Chen et al., 2020).

The usability of these applications is a critical factor that directly impacts productivity and tool adoption. An inefficient, unclear, or non-intuitive experience leads to decreased productivity and abandonment of the system by users (Nielsen, 2012). Perceived ease of use, in turn, directly influences the acceptance and effective use of technology (Davis, 1989; Venkatesh et al., 2003). Usability is defined by ISO 9241-11 (2018) and by Nielsen (1993) as the extent to which a product can be used by specified users to achieve objectives with effectiveness, efficiency, and satisfaction. User experience (UX), according to Hassenzahl and Tractinsky (2006), is a determinant for the adoption and continued use of digital technologies, with recent perspectives on corporate applications (Kim; Kim; Lee, 2022; Ali; Khalid; Sheraz, 2023).

Despite the growing investment and use of corporate applications for salespeople, there is a gap in understanding the behavioral patterns of these users. Much of the usability research still focuses on the end consumer, not on the internal corporate user, such as sales teams (Ghezzi et al., 2018; Venkatesh et al., 2003). This theoretical deficiency prevents a deep understanding of how salespeople use these platforms in their real work contexts, hindering strategic decisions related to interface improvements, training, and feature customization.

To fill this gap, it is essential to understand real user behavior, which is indispensable for the development of effective experiences (Nielsen; Loranger, 2006). Usability measurement allows for the identification of barriers and guidance for design corrections (Sauro; Lewis, 2016). An effective approach for this is user segmentation based on objective interaction patterns recorded in application usage logs. These logs store events such as screen accesses, clicks, and feature executions, allowing the extraction of engagement and usability metrics to characterize behavior profiles at scale (Nielsen; Loranger, 2006; Sauro; Lewis, 2016).

Among segmentation techniques, clustering algorithms, such as K-means, stand out for their ability to group observations based on similarity between variables (Tan; Steinbach; Kumar, 2019). K-means is widely used due to its computational efficiency and interpretability (Jain, 2010; MacQueen, 1967), partitioning data into groups by minimizing distance to the centroid (Han; Kamber; Pei, 2012). Thus, usability analysis combined with the application of clustering algorithms, such as K-means, becomes an essential resource for directing continuous improvements in the application and maximizing the impact of technology use on business performance.

Understanding how salespeople use the application, segmenting them based on real behavior patterns, contributes to generating objective insights about the user experience, prioritizing platform evolution needs, and supporting data-driven strategic decisions. Given the above, this work aims to segment users of a sales application based on the use of key screens and functionalities, using the K-means algorithm.

2. Material and Methods

This study was characterized as a quantitative and applied research, configuring itself as a case study. The central objective was to segment users of a sales application based on the use of key screens and functionalities, employing the K-means clustering methodology. The research sought to understand different usage profiles to support strategic decisions related to product improvements, using objective behavioral data.

The data used were obtained from application usage records, corresponding to interactions carried out by sellers of a multinational company. The analyzed sample comprised 21,878 sellers, selected from a larger user base, located in four countries: Brazil, Mexico, Colombia, and Argentina. Data collection covered the period from January to December 2025, focusing on the most relevant markets for the application.

The data collection was of a secondary nature, based exclusively on system usage logs, and did not involve direct instruments such as interviews or questionnaires. All data were rigorously anonymized, ensuring confidentiality and ethical compliance, without any identification of users or companies. Information extracted from the logs included application access frequency, interaction volume per session, usage of key functionalities, number of events executed, and usage recurrence and consistency.

For segmentation, variables reflecting user usability and engagement were selected, based on classic concepts from the literature on user experience (UX) and application usability (Nielsen, 1993; ISO 9241-11, 2018). The metrics were categorized into Efficiency, Effectiveness, and Satisfaction and Engagement. Efficiency measured the speed and resources used, including average number of clicks per task, average time to complete an order, screens accessed per interaction, and events per session.

Efficacy assessed the precision and completeness with which users achieved their objectives, considering the success rate of orders, error frequency, interactions required to complete sales operations, and average use of key functionalities. Satisfaction and Engagement measured user comfort, acceptance, and involvement, through metrics such as recurrence and consistency of use, session frequency, and interaction with recommended functionalities. In total, 25 distinct events were used, classified into three categories: direct action, indirect action, and information visualization.

The data preparation and cleaning stage was crucial to ensure they were consistent and comparable. Data was aggregated by user, creating metrics that represent individual behavior, such as the total number of events and clicks, the average clicks per event, the quantity of distinct events performed, the quantity of distinct feature types used, and the proportion of active months in the analyzed period. Missing values were non-existent, and all variables were converted to the numeric type.

Next, MinMax normalization was applied to transform each metric to the range of 0 to 1. This transformation was essential for all variables to have equivalent weight in identifying user groups, preventing scale differences between metrics from distorting cluster formation, as the K-means algorithm is sensitive to data magnitude.

The data analysis was conducted using the K-means clustering algorithm, an unsupervised learning technique. K-means partitions the data into k clusters, grouping observations based on the similarity between variables (MacQueen, 1967; Jain, 2010). The algorithm’s operation was based on minimizing intra-cluster variance, measured by the sum of squared distances between each observation and the centroid of its respective cluster.

Initially, the algorithm defined k centroids, and each user was assigned to the cluster whose centroid presented the smallest Euclidean distance. Then, the centroids were recalculated as the mean of the observations assigned to each cluster. This iterative process was repeated until the algorithm reached convergence, meaning when there were no relevant changes in the composition of the clusters. The analyses were performed using the Python language, with specialized libraries for data science and machine learning.

To determine the most appropriate number of clusters (k), the Elbow method and Silhouette coefficient analysis were employed (Thorndike, 1953; Rousseeuw, 1987). The Elbow method analyzed the sum of squared distances within clusters (inertia), seeking an inflection point in the curve that indicated a balance between model complexity and data representation. Complementarily, the Silhouette coefficient assessed internal cohesion and separation between clusters, with higher values indicating well-defined groupings. The number of clusters was defined based on the joint analysis of these methods.

3. Results and Discussion

This section details the results obtained from applying the K-means clustering methodology to segment users of a sales application, as per the central objective of this study. The findings reveal distinct usage patterns, characterized by engagement and usability metrics, which provide insights into seller interaction with the platform. Cluster analysis allowed for the identification of user profiles, differentiated by the intensity, depth, and diversity of interactions, and not just by access frequency, which is crucial for understanding usability in corporate contexts. The results are presented in stages, from data characterization to profile interpretation.

Initially, the database used for the research comprised anonymized usage records from a sales application, collected from 21,878 salespeople across four countries: Brazil, Mexico, Colombia, and Argentina. The analysis period spanned from January to December 2025, focusing on the most relevant markets for the application. Recorded interactions included events, clicks, and feature usage, categorized into three main types: indirect action, direct action, and information viewing. Indirect action referred to viewing assigned tasks and recommended products, while direct action involved concrete sales operations, such as order registration. Information viewing encompassed access to customer data, order history, and performance indicators, all essential for supporting the salesperson during negotiation.

For the segmentation, variables reflecting user usability and engagement were selected, aligned with Nielsen’s (1993) principles and the ISO 9241-11 (2018) standard. The metrics were grouped into three categories: efficiency, which assesses speed and resources used; effectiveness, which measures accuracy and completeness in achieving objectives; and satisfaction and engagement, which gauge user comfort and involvement. These metrics included the total number of events, click volume, average clicks per event, number of distinct events, diversity of utilized functionality types, and regularity of use, providing a comprehensive view of individual behavior.

The descriptive analysis of the usage variables revealed that the total number of events presented a mean of 18.89, with a median of 21.0 and a standard deviation of 6.59. The total number of clicks registered a mean of 3,089.19, a median of 1,603.0, and a standard deviation of 4,180.85, indicating high dispersion. The mean clicks per event was 135.69, with a median of 79.14 and a standard deviation of 171.91. Distinct events had a mean of 18.18, a median of 20.0, and a standard deviation of 6.23. The distinct types of functionalities accessed presented a mean of 2.43, a median of 3.0, and a standard deviation of 0.76. Usage regularity, in turn, had a mean of 1.52, a median of 1.67, and a standard deviation of 0.52. The metrics related to the number of clicks exhibited the highest dispersion, which highlighted the need for data normalization for the application of K-means.

The preparation and normalization of data were crucial steps to ensure the consistency and comparability of variables, avoiding distortions in cluster formation. The original records, which corresponded to individual interactions, were aggregated by user to create metrics representing individual behavior. Metrics for total events, total clicks, average clicks per event, distinct events, distinct types, and regularity of use were calculated from the actual records. No missing values were identified, and all variables were converted to numeric type. MinMax normalization was applied, transforming each metric to the range of zero to one, ensuring that all variables had equivalent weight in identifying user groups, as recommended by Nielsen (1993) and ISO (2018) for behavioral analysis.

The application of the K-means algorithm required defining the ideal number of clusters (k). For this purpose, the Elbow method and Silhouette coefficient analysis were employed, techniques widely used in the literature for evaluating the quality of groupings in unsupervised clustering models (Jain, 2010). The Elbow method, which analyzes the sum of squared distances within clusters (inertia), indicated an inflection point at k = 3, suggesting this value as an appropriate choice, as the reduction in inertia becomes less significant from this point onwards (Thorndike, 1953). This point represents a balance between model complexity and data representation capability.

Complementarily, the analysis of the average Silhouette coefficient, which assesses internal cohesion and separation between clusters (Rousseeuw, 1987), showed that, although lower values of k presented slightly higher coefficients, adopting three clusters provided a greater capacity for differentiation and interpretation of the analyzed profiles, without significantly compromising the quality of the groupings. The joint analysis of both methods confirmed that the value k = 3 offers an adequate balance between model complexity and segmentation quality. Thus, the value k = 3 was adopted as the final number of clusters for this study, allowing for the identification of distinct levels of engagement and usability (low, medium, and high).

The clustering resulted in the identification of three distinct user profiles, named Occasional Users (Cluster 0), Regular Users (Cluster 1), and Advanced Users (Cluster 2). The differences between these groups were significant, mainly in the metrics related to the total number of clicks, the average clicks per event, and the diversity of events and types of functionalities accessed. The usage regularity variable, although included in the model, showed low variability among the clusters, indicating that it was not a discriminating factor for the segmentation of profiles, which were primarily defined by the other five variables.

Cluster 0, named Occasional Users, presented the lowest engagement level, with an average of 8.6 events, 251 total clicks, and 21.7 clicks per event. This group performed approximately 8 distinct events and accessed, on average, 2.4 types of functionalities, suggesting a punctual and limited use of the application. Although the regularity of access was similar to other clusters, the low intensity of interaction and restricted exploration of functionalities characterize this profile as users who perform simple and objective tasks, without delving into the platform’s resources.

Cluster 1, identified as Regular Users, demonstrated an intermediate level of engagement. This group registered an average of 21 events, 2,252 total clicks, and 101 clicks per event. Additionally, they accessed approximately 20 distinct events and 3.73 types of functionalities. These users interact with the system more deeply, exploring various functionalities and performing more complex tasks, albeit still at a moderate level compared to the highest engagement group. The regularity of use also proved similar to the other clusters, reinforcing the idea that the difference lies in the quality and depth of interaction.

Finally, Cluster 2, named Advanced Users, represented the most active and engaged group. With an average of 24 events, 10,225 total clicks, and 425 clicks per event, this profile stood out for its high volume and intensity of interactions. Advanced Users registered 23 distinct events and 3.68 types of functionalities accessed, indicating that they explore most of the application’s resources and consistently perform complex tasks. This behavior suggests experienced or system-dependent users, who intensely use its functionalities and interact in-depth, even with a similar regularity of use to the other groups.

The findings reinforce the idea that access frequency, or regularity of use, is not sufficient to explain an application’s usage maturity, as pointed out by Sauro and Lewis (2016). The substantial differences between the clusters in efficiency and effectiveness metrics, such as total number of events, click volume, average clicks per event, and diversity of accessed functionalities, indicate that users are fundamentally distinguished by how they use the system, not just by “how much” they use it. This differentiation is crucial for understanding perceived usability and usage maturity as central factors in segmentation.

The application of the K-means algorithm proved to be viable and effective for identifying consistent behavior patterns in a sales application environment (Jain, 2010; Tan; Steinbach; Kumar, 2019). The identified profiles provide a solid foundation for strategic decisions and system improvements, allowing for the prioritization of functionalities, optimization of usage flows, and targeting of specific actions to increase the engagement and satisfaction of each user type, aligning with Nielsen’s (1993) and ISO 9241-11 (2018) usability principles. This approach contributes to the personalization of engagement strategies, design optimization, and data-driven management.

The main differential of this study lies in the integration of usability, engagement, and real usage behavior metrics, extracted directly from system logs, in a large-scale business context with high operational complexity. Unlike many studies that rely on self-reported metrics, this research demonstrates how classic usability principles can be operationalized through objective observational data. The evidence that users with similar usage frequency can exhibit radically distinct interaction profiles reinforces the need for more granular and behavioral approaches to segmentation in corporate systems, going beyond sociodemographic variables or aggregated access metrics.

From a practical point of view, the results offer relevant contributions to companies that use digital applications to support commercial operations. The identification of distinct profiles allows for the development of specific actions for each group, such as targeted training for occasional users or advanced functionalities for experienced users. Understanding the functionalities explored by advanced users can guide usability improvements and workflow simplification for less engaged users. Furthermore, behavioral segmentation provides objective subsidies for decision-making, allowing for monitoring user evolution between clusters over time and personalizing product improvements to optimize the performance of each group. This maximizes the return on investment in technology, focusing on the quality of use and encouraging behaviors that contribute to commercial performance and user satisfaction.

In summary, this study demonstrated that the joint analysis of engagement and usability metrics, combined with clustering techniques, constitutes a robust approach to understanding user behavior in corporate systems. By highlighting that regularity of use is not sufficient to explain behavioral differences, the research reinforces the importance of investigating the depth, diversity, and quality of interactions, offering relevant theoretical and practical contributions to the field of information systems, usability, and data-driven commercial management, fulfilling the objective of segmenting users based on the utilization of key screens and functionalities of the application.

4. Conclusion

The study aimed to segment users of a sales application based on the utilization of key screens and functionalities, employing the K-means algorithm. The existence of three distinct user profiles – occasional, regular, and advanced – was verified, identified from engagement and usability metrics extracted from system usage logs. The findings revealed that the differentiation between these groups occurred mainly due to the intensity, depth, and diversity of interactions with the platform, rather than access regularity. It was observed that usage frequency, although present, was not the discriminating factor for segmentation, indicating that users with similar access patterns may exhibit radically different interaction behaviors in terms of feature exploration and action volume. This distinction reinforces the understanding that the quality and depth of use are more relevant than mere frequency in characterizing the interaction maturity with corporate systems.

The main contribution of this work lies in the operationalization of classic usability principles through objective observational data, providing subsidies for the personalization of engagement strategies, design optimization, and data-driven management in large-scale commercial environments. The identification of these profiles allows companies to develop specific actions, such as targeted training for occasional users or improvement of functionalities for advanced users, maximizing the return on investment in technology. However, the study, quantitative and case-based in nature, did not explore differences in usage by country, delve into the reasons for the high click volume of advanced users, nor analyze the temporal evolution of users among clusters or the direct correlation of profiles with commercial performance. It is therefore suggested that future studies investigate these gaps, including comparative analyses by country, qualitative research on click behavior, and longitudinal studies to map migration between profiles and the relationship with sales results.

Bibliographic References

ALI, M.; KHALID, S.; SHERAZ, M. Usability evaluation of mobile business applications. Journal of Systems and Information Technology, v. 25, n. 2, 2023.

BHARADWAJ, A.; EL SAWY, O. A.; PAVLOU, P. A.; VENKATRAMAN, N. Digital Business Strategy: Toward a Next Generation of Insights. MIS Quarterly, 2013.

CHEN, J. et al. Impact of Mobile Sales Applications on Employee Performance. Journal of Business Research, 2020.

DAVIS, F. D. Perceived Usefulness, Perceived Ease of Use, and User Acceptance of Information Technology. MIS Quarterly, v. 13, n. 3, p. 319–340, 1989.

GARTNER. Forecast: Mobile Applications Worldwide, 2021–2025. Gartner Research, 2021.

GHEZZI, A. et al. Digital transformation and business model innovation: Role of mobile applications. Technological Forecasting & Social Change, 2018.

HAN, J.; KAMBER, M.; PEI, J. Data Mining: Concepts and Techniques. 3. ed. Morgan Kaufmann, 2012.

HASSENZAHL, M.; TRACTINSKY, N. User experience – A research agenda. Behaviour & Information Technology, v. 25, n. 2, p. 91–97, 2006.

INTERNATIONAL ORGANIZATION FOR STANDARDIZATION. Ergonomics of human-system interaction – Part 11: Usability: Definitions and concepts. ISO 9241-11:2018. Geneva: ISO, 2018.

JAIN, A. K. Data clustering: 50 years beyond K-Means. Pattern Recognition Letters, v. 31, n. 8, p. 651-666, 2010.

KIM, J.; KIM, J.; LEE, Y. Understanding user experience in mobile enterprise applications. International Journal of Human-Computer Studies, v. 158, 2022.

MACQUEEN, J. Some methods for classification and analysis of multivariate observations. In: BERKELEY SYMPOSIUM ON MATHEMATICAL STATISTICS AND PROBABILITY, 5., 1967, Berkeley. Proceedings… Berkeley: University of California Press, 1967. p. 281-297.

NIELSEN, J. Usability 101: Introduction to Usability. Nielsen Norman Group, 2012.

NIELSEN, J.; LORANGER, H. Prioritizing Web Usability. Berkeley: New Riders, 2006.

NIELSEN, Jakob. Usability Engineering. San Francisco: Morgan Kaufmann, 1993.

OLANREWAJU, R. et al. Mobile Application Usage Patterns: Implications for Service Design. International Journal of Mobile Human Computer Interaction, 2020.

ROUSSEEUW, P. J. Silhouettes: a graphical aid to the interpretation and validation of cluster analysis. Journal of Computational and Applied Mathematics, v. 20, p. 53–65, 1987.

SAURO, J.; LEWIS, J. R. Quantifying the User Experience: Practical Statistics for User Research. 2. ed. Amsterdam: Morgan Kaufmann, 2016.

TAN, P.-N.; STEINBACH, M.; KUMAR, V. Introduction to Data Mining. 2. ed. Pearson, 2019.

THORNDIKE, R. L. Who belongs in the family? Psychometrika, v. 18, n. 4, p. 267–276, 1953.

VENKATESH, V.; MORRIS, M. G.; DAVIS, G. B.; DAVIS, F. D. User Acceptance of Information Technology: Toward a Unified View. MIS Quarterly, v. 27, n. 3, p. 425–478, 2003.

VENKATESH, V.; THONG, J. Y. L.; XU, X. Consumer Acceptance and Use of Information Technology: Extending the Unified Theory of Acceptance and Use of Technology. MIS Quarterly, 2012.

Article originating from the Final Course Project of the Specialization in Data Science and Analytics of the MBA USP/Esalq

To learn more about the course, click here and access the MBX Academy platform