Predicting Termination of Mobile Subscribers Using Classification Algorithms
No Thumbnail Available
Date
2019-12
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Addis Ababa University
Abstract
Service termination is one key challenge of operators that can significantly reduce their
revenue. It is still a challenge even in Ethiopian monopoly market even if users do not
have alternative service provider. For instance, the market has seen 1.6 percent overall
prepaid mobile subscriber termination rate within a quarter, let alone specific service
termination. To address this challenge, operators need to understand causes of
termination and take timely proactive actions to mitigate number of terminations. For
this purpose, they need to accurately and timely predict subscribers with potential service
termination based on collected service related data. Performance of various classification
algorithms have been investigated to predict subscribers’ behavior. However,
performance of such algorithms are not studied in the context of Ethiopia where ethio
telecom has an enormous data that helps to define its subscribers’ behavior.
Objective of this thesis work is to study performance of classification learning algorithms
for predicting termination in the context of Ethiopian prepaid subscribers. Studied
algorithms are J48, Random Forest and Naïve Bayes and their performance are compared
in terms of prediction accuracy, performance, errors and interpretability of their model.
Algorithms effectiveness and efficiency are evaluated considering two validation
methods (Percentage Split and Cross Validation) and three data sets. For the performance
evaluation, we used WEKA 3.8 tool algorithm implementation.
Obtained results show that Random Forest scores the highest prediction accuracy while
Naïve Bayes scores the least. Random Forest and Naïve Bayes scores their best at 93.4% and
85.9% respectively, besides J48 scores 93.3%. J48 is as accurate & robust as Random Forest.
Moreover, it provides the most interpretable and clear model.
Description
Keywords
J48, Random Forests, Naïve Bayes, service termination, subscribers’ defection, churn, confusion matrix