Recommendations on Recommendations: Predicting Content Diffusion and Engagement in Online Social Networks

Abstract

This paper explores how content characteristics and users' public profiles can be used to predict information diffusion and engagement in online social networks. In order to improve on past diffusion research, first a new dataset was collected on 152 Facebook users using a Chrome extension. The extension collected users' profile information, and tracked the articles that users saw on their News Feeds and which ones they clicked on. In total, 11,787 unique user-article pairs were collected. Using this dataset, link prediction methods and a diffusion model that incorporates user and content features were compared. The diffusion model achieved an accuracy of 85.8% and outperformed the link prediction methods, indicating that user and content features are important for predicting diffusion. To explore content engagement, two methods were used: a novel engagement model and a random forest classifier. To predict whether or not a user would engage with a given piece of content, the model quantified important characteristics like the user's similarity with the content, the user's trust in the source of the content, and the similarity between the content and all past content they have interacted with. Both the model and the random forest classifier had high accuracy rates of 85.4% and 85.1% respectively. The user's relationship with the information source and the presence of emotionally-charged words in the title and comments of the article were found to be very important features for the random forest classifier.