Forecasting Retweet Count during Elections Using Graph Convolution Neural Networks
Abstract
A retweet refers to sharing a tweet posted by another user on Twitter and is a primary way information spreads on the Twitter network. Political parties use Twitter extensively as a part of their campaign to promote their presence, announce their propaganda, and at times debating with opponents. In this work we consider the problem of early prediction of the final retweet count using information from the network during the first several minutes after a post is made. Such predictions are useful for ranking and promoting posts and also can be used in combination with fake news detection. From a machine learning perspective, the task can be viewed as a regression problem. We introduce a novel graph convolution neural network for forecasting retweet count that combines network level features through a graph convolution layer as well as tweet level features at a higher dense layer in the network. We first will provide an overview of the graph convolution network architecture and then perform several experiments on Twitter data collected during presidential elections in South Africa, Kenya, and Nigeria. We show that the model outperforms baseline models including a feed-forward neural network and the popular point process based model SEISMIC.
Degree
M.S.
Advisors
Mohler, Purdue University.
Subject Area
Artificial intelligence|Computer science
Off-Campus Purdue Users:
To access this dissertation, please log in to our
proxy server.