Developing a Predictive Model for Message Propagation on Online Social Networks

Abstract

We studied the problem of predicting the growth of retweet cascades over Twitter from structural and temporal points of view. We rst devised a new measure of cascade growth that overcomes the skew problem and, at the same time, takes into consideration the statistical characteristics of the dataset. We suggested a starter list of Twitter features suitable for cascade growth prediction. Then, we proposed a hybrid feature selection approach to select the most discriminating features based on the dataset. We presented and analysed the prediction accuracy results using either all features or the reduced set of features. Two prediction models were built using Random Forest and Multilayer Perceptron.