A Comparative Evaluation of SARIMA and LSTM Models for Facebook Interaction Forecasting

Abstract

The massive amount of data generated daily by social networks such as Facebook offers valuable opportunities for understanding user behavior, but also poses significant challenges for prediction. This paper presents a comparative study of two forecasting approaches for mining Facebook interactions time series: the classical SARIMA model and the LSTM neural network. Using eight years of monthly data, the final comparison reveals that LSTM significantly outperforms SARIMA across all evaluation metrics. Specifically, LSTM achieves a 38.9% lower RMSE, a 37.4% lower MAPE, and an accuracy of 87.6% compared to 80.2% for SARIMA. Furthermore, LSTM demonstrates near-zero prediction bias, while SARIMA shows systematic overestimation. These results confirm that for social networks data mining, LSTM provides more reliable and accurate forecasts, particularly for medium- to long-term planning applications.