Machine Learning Based Mobile Airtime Credit Risk Prediction using Customer Profile and Loan Information
No Thumbnail Available
Date
2021-12
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Addis Ababa University
Abstract
Airtime credit service is usually used by prepaid mobile subscribes when they cannot recharge their service number. The credit service helps to increase service subscribers' satisfaction and revenue of the telecom operator. However, the service has its own risk as there is no guarantee that the prepaid subscribers would pay back the credit they took. To solve this issue, several researchers propose to predict airtime credit risk using machine learning based approaches. To build the prediction models, these approaches use either customer or customer activity related information as features. These features cover only part of the customer related information captured by telecom companies. Other potentially relevant information such as education of the customer, subscriber loan amount and frequency, and recharge frequency are not taken into consideration to predict airtime credit risk. In this research work, we propose an approach that takes in to consideration education and recharge frequency from customer profile and subscriber loan amount and loan frequency from the loan information to predict airtime credit risk.To assess the impact of the proposed approach, we conducted an experiment using 90,000 mobile subscriber’s data. The experiment uses four machine learning algorithms: decision tree (DT), logistic regression (LR), random forest (RF) and (MLp). The results show that the combination of the existing and new features improves airtime credit risk prediction for all algorithms. The highest improvement, i.e., 6.91%, in accuracy while using the existing customer profile and usage, and new features is observed by LR. For the existing loan and new features, the highest improvement, i.e., 7.8%, in F-measure is observed by RF. We ranked the features based on their importance using feature ranking algorithms. Subscriber loan amount is the most important feature. Loan frequency and recharge frequency are also in the top seven features that contributed to the risk prediction. This shows that the newly added features have helped to enhance the airtime credit risk prediction.
Description
Keywords
Airtime credit risk, Airtime credit risk prediction, Machine Learning Algorithms