Customer Segmentation On Mobile Money Users In Kenya

Abstract

Customer segmentation enables organizations to partition a market into subsets that have common needs, interests and priorities. This helps businesses to come up with design and strategies that fulfills the customer needs. Usage of mobile money services has been widely adopted in Kenya with over 22 million customers using the service. Mobile money operators are mainly telecommunication companies leveraging on their infrastructure and customer base. Many businesses and individuals in Kenya use mobile money for services such as peer to peer transfers, payment services and financial services such as mobile banking. Previous segmentation of mobile money customers has mostly been done basing on the types of transaction or demographic factors as in banking such as age, gender, assets and location however this does not give a 360 degrees overview of the customer enabling an opportunity for improvement. Mobile network operators own big volumes of data which can enable improvements on segmentation models e.g subscriber’s network activity i.e. calls, sms, data usage, demographic factors such as age and behavioral factors such as types and frequency of loans and payments. Leveraging on this data to identify different customer groups and their needs is important for service providers to respond to changing customer demands, cope with fast technological advancements and innovate around local market conditions. Using network and mobile money data, this study compares various clustering algorithms aiming at identifying the algorithm that creates the most solid customer profiles. Hierarchical clustering, KMeans and affinity propagation algorithms were used to segment customers and compared using internal validation measures. Our dataset comprised of various demographic and behavioral features obtained from a telecommunications company data warehouse. Co-relation between the features was tested enabling us to focus on age, network revenue, amounts transacted on mobile money, frequency of loan uptake, customer and organization transfers, goods and service payments and deposits and withdrawals as our features for modelling. The dataset was then fit into our algorithms. Agglomerative clustering generated seven clusters with a normalized mutual score of 0.5526 and adjusted rand score of 0.5436 and silhouette coefficient of 0.4523. KMeans generated 11 clusters with an NMI score of 0.5168 and adjusted rand score of 0.3315 and silhouette coefficient of 0.2369. Affinity propagation generated the largest number of clusters of 504, had a memory utilization of 91% and took the longest time to execute. This established AP as unsuitable for our dataset. Agglomerative clustering had the best performance in terms of the compactness and connectedness of clusters however clusters obtained from KMeans were more granular as compared to agglomerative clustering segments.