PREDICTING THE TYPE OF ENGAGEMENT FOR UNIVERSITIES' TWITTER FEEDS

Abstract

Many studies show how to engage with audiences on social media, but a lack of studies shows how universities use social media accounts in the scientific research domain. Therefore, based on the research gap, the present study aims to contribute to the field of predicting the most probable type of engagement (like, retweet, or reply) for ten university official Twitter accounts. Moreover, the study also proposes to find some of the features contributing to this prediction. In order to predict the type of interaction, the research uses a combination of human- selected and machine-extracted features to train three machine learning models (Logistic Regression, Random Classifier, and LightGBM) and a deep learning model (neural network using BERT model). Human selected features are mainly binary variables that contain tweet information, while machine-extracted features are large-dimensional features that we obtain from the texts of the tweets. The results show that by combining both types of features, we can predict the most probable type of engagement and an overview of the features that contribute to this prediction, such as if the tweet contains a hashtag or if the tweet is a reply. Also, the findings show that the best method to predict this engagement is LightGBM and neural networks. Research and practical implications include helping practitioners to create the content strategy based on the engagement objectives and providing more knowledge to help them understand which features contribute to the type of engagement.