Forecasting Facebook User Engagement using Hybrid Prophet LSTM and iForest

Abstract

Business forecasting remains a popular topic these days. A reliable business forecast often plays a vital role in an advertising campaign. The amount of attention acquired by posting an advertisement is one of the most essential criteria for determining the effectiveness of the advertisement. The number of times that public users engage with a content signifies the amount of attention received, was measured by user engagement.  With a good forecast, the advertisement could be promoted to a larger number of people. Facebook, as the most popular social media site, is preferred by the majority of advertisers. Therefore, this study addresses Facebook user engagement by forecasting the optimum date to post an advertisement.  Different forecasting models, each with its own strengths and weaknesses, are used to model time series data with various properties. The objective of this study is threefold: to investigate the accuracy of the proposed Hybrid Prophet-LSTM that combines Long Short-Term Memory (LSTM) and FBProphet (Prophet), to study the holiday impact on user engagement forecasting on Facebook brand pages, and to study the effect of implementing iForest on the dataset and its contribution to the forecast result. Data from three popular brand pages were used in the experiments in the period of June 2018 to March 2021. The results show that the proposed hybrid model outperforms both the standalone LSTM and Prophet across the datasets. Besides, it is found that holiday effect could generally increase forecast accuracy. In general, datasets pre-processed using iForest can reduce the forecast error under specific conditions. Therefore, the optimum date for an advertisement campaign can be determined on the basis of the most anticipated user engagement, which consequently enhances the business income.