Desarrollo de un modelo basado en machine learning para pronosticar el número de usuarios de televisión por suscripción para el departamento de Cundinamarca

Abstract

In Colombia, the landscape of subscription television has undergone significant changes due to increasing competition from streaming platforms and the country's geographical diversity. This phenomenon highlights the need for predictive models that anticipate subscriber behavior. In response, this monograph-based research project proposes the development of a machine learning model to forecast the number of subscription television users in the department of Cundinamarca. The methodology centers on the use of Meta’s (Facebook) Prophet library, renowned for its effectiveness in modeling time series with seasonal patterns and long-term trends. Prophet enables flexible adjustments of trend and seasonality components, making it particularly suitable for telecommunications data, where recurring variations over time are common. The choice of this tool is based on its ease of implementation, ability to handle missing data, and robustness against sudden changes. To build the model, historical data from the “Subscription TV Subscribers” Dashboard of the Communications Regulation Commission (CRC) was utilized. This dataset provides disaggregated information based on technology types (HFC Digital, Satellite, IPTV, among others) and geographical zones. Through exploratory data analysis (EDA), variable normalization, and feature selection, the dataset was prepared to feed the predictive model. The results not only provide projections for new subscriber numbers but also identify significant patterns in service subscription behavior. Key influencing factors include technology type, socioeconomic strata, and the geographical location of users. Additionally, the Prophet-based predictive model demonstrated solid performance, reflected in metrics such as Mean Absolute Error (MAE) and Root Mean Square Error (RMSE), successfully forecasting subscription television user trends for the next 6 to 12 months. In conclusion, the development of this model serves as a strategic tool for telecommunications operators and establishes a robust foundation for public policy design focused on digital inclusion and improved access to ICT services in Cundinamarca. Future research directions could integrate additional socioeconomic variables or explore hybrid models combining Prophet with Recurrent Neural Networks (RNNs) to further refine prediction accuracy. Let me know if you need refinements!