Top-Up Forecasting of Pre-Paid Mobile Subscribers

• Autor(a) principal: Alves, Pedro Miguel Ferreira
• Data de Publicação: 2021
• Tipo de documento: Dissertação
• Idioma: eng
• Título da fonte: Repositório Científico de Acesso Aberto de Portugal (Repositórios Cientìficos)
• Texto Completo: http://hdl.handle.net/10400.22/20157
• topic: Telecommunications; Top-up; Forecasting; Big Data; Machine Learning; Telecomunicações; Recarga; Previsão

In an ever-evolving technology world, telecommunications operators must attend to client needs in an effective and speedy manner to strengthen their relationship. The difficulty of this challenge is heightened in Big Data environments where there is a necessity to make sense of the valuable information within data. In the pre-paid telco environment, also known as pay-as-you-go, it is imperative for operators to predict client behaviour efficiently to meet their needs and improve campaigns and notifications, thus improving communication, client retention and revenue. In this dissertation, a novel top-up date and value prediction solution for the prepaid telco environment, is presented. This solution aims to dynamically estimate, for each client, the top-up date and value for the upcoming month. For this, the initial data goes through the developed processing pipeline. The first step is pre-processing, where data is cleaned and transformed. After this, it undergoes a feature engineering and selection step to identify the most relevant features for the prediction of the monthly frequency and value. For the prediction of the targets, several regression techniques were studied both on the offline and online scenario with the help of sliding windows. Using the most efficient technique, the monthly target predictions undergo a processing stage in which they are transformed into the individual top-up date range and top-up monetary value range for the following month. The evaluation of these predicted ranges is based on verifying if the observed event falls within the predicted interval. The solution is implemented in Python and the Jupyter Notebooks environment for data analysis, dimensionality reduction and offline learning experiments. The online learning experiments make use of the Massive Online Analysis (MOA) graphical user interface (GUI) framework. In the end, the designed solution is able to estimate individual top-up activity with an accuracy of approximately 80 % for the date and 70 % for the monetary value.