Browsing by Author "Demelash, Biru"

Now showing 1 - 1 of 1
  • No Thumbnail Available
    Item
    Usage Based Clustering of Customers for Mobile Service Packaging
    (Addis Ababa University, 2019-12) Demelash, Biru; Beneyam, Berehanu (PhD)
    Satisfaction of customers is the most important factor for mobile operators to be successful. This needs effective customer segmentation and segment targeted mobile service packaging and delivery. Segmentation differentiates customers into multiple groups that manifest different service needs and preferences, thus different service packages. It has been traditionally performed using demographic and value-based segmentation methods based on customer survey data. For improved efficiency, advanced clustering techniques that exploit existing historical customer data from network management system have been applied. Instead of using a single dimension of value-based segmentation, the historical data set with many features was applied to assess the customer service usage behavior from different dimensions. For a dataset with many attributes, such advanced clustering techniques have not been investigated in the Ethiopian context. The thesis work investigates and compares the performance of K-means and expectation-maximization algorithms for usage-based clustering using voice, SMS and internet service usage call detail record data of mobile customers. The performance was compared using metrics such as cluster size or ratio, cluster cohesion or compactness and separation between centroid values. These metrics were used to evaluate the quality of the clustering result of the algorithms in identifying distinguished customer segments from each service usage dataset for mobile service packaging purposes. Optimal cluster size per dataset was determined using elbow method. In the study, data processing and algorithm implementations were performed using WEKA data mining tool. Achieved results indicate that for all the datasets the EM algorithm formed compact clusters with low level of within cluster variance. On the other hand, K-means clustering has a better quality in assigning instances to each cluster fairly. In general, the study identified important additional attributes from the CDR dataset to differentiate customers for mobile service packaging purpose. These additional features enhance the insight on customers to provide well differentiated mobile service packages.