Link prediction on tweets’ content

Abstract

In this paper we test various weighted local similarity network measures for predicting the future content of tweets. Our aim is to determine the most suitable measure for predicting new content in tweets and subsequently explore the spreading positively and negatively oriented content on Twitter. The tweets in the English language were collected via the Twitter API depending on their content. That is, we searched for the tweets containing specific predefined keywords from different domains - positive or negative. From the gathered tweets the weighted complex network of words is formed, where nodes represent words and a link between two nodes exists if these two words co-occur in the same tweet, while the weight denotes the co-occurrence frequency. For the link prediction task we study five local similarity network measures commonly used in unweighted networks (Common Neighbors, Jaccard Coefficient, Preferential Attachment, Adamic Adar and Resource Allocation Index) which we have adapted to weighted networks. Finally, we evaluated all the modified measures in terms of the precision of predicted links. The obtained results suggest that the Weighted Resource Allocation Index has the best potential for the prediction of content in tweets.