Towards Feature Selection for Cascade Growth Prediction on Twitter

On online social networks such as Twitter, retweeting allows users to share a variety of content to their own followers. As tweets are retweeted from user to user, large cascades of tweets propagation are formed. During the past decade, social networks have been used for political rallying, civil society campaigns, and marketing promotions. The growth of cascades over time signal the popularity or lack thereof of the subject matter. In this work, we pose the question of whether the same feature set can be used for cascade growth prediction of any dataset on Twitter. We find evidence that features governing the cascade growth vary from one dataset to another. We first devise a definition of structural and temporal growth. Then, we propose an approach to select the best of these features based on the dataset for better accuracy results. We examine two types of growth prediction: structural and temporal. We use both Random Forest and Multilayer Perceptron as our models for growth prediction on data sets of two political campaigns in Egypt. The campaigns were concurrent in time, and were rallying for opposite causes, with a noticeable difference in societal popularity. We present and discuss the results of the most discriminating features in predicting cascades' growth of the two data sets, and provide evidence that the preselection of features improved the accuracy of the prediction task on the two data sets studied.