Comparison of Prediction Methods on Large-Scale and Long-Term Online Live Streaming Data

Effective prediction of online live streaming traffic plays a crucial role not only in optimizing network resource allocation for enhancing viewer experience but also in assessing factors impacting audience retention and the overall sustainability of streaming platforms. This study conducts a comprehensive evaluation of machine learning methods for online live streaming traffic prediction using extensive hourly traffic data. The dataset comprises 1,385,444,808 live streaming entries and encompasses 30,690,841 unique streamers from the Douyu platform, spanning December 2020 to April 2023. Various experimental settings are employed to compare the performance of these methods. Our findings reveal that among ten methodologies considered, the Bidirectional Long Short-Term Memory, Extra Tree (ET), and Random Forest models demonstrate consistent and robust performance. Particularly, the ET model exhibits outstanding accuracy and precision in predicting daily viewer counts when incorporating pertinent features. In the domain of large-scale and long-term live streaming data prediction, machine learning approaches surpass traditional time series forecasting methods. Moreover, our analysis underscores the significance of incorporating streamer count in enhancing the accuracy of network traffic prediction. Interestingly, while hourly features show limited impact, in certain scenarios, their inclusion may even diminish the predictive efficacy of the models.