Controversy prediction in Twitter

Abstract

Controversy is a complex subject that has attracted the attention of research work in different fields. In social media, the detection of controversy is a big challenge due to the huge amount of information that is expressed by large audiences, containing opinions for news, events and any kind of stimulation. The current work focuses on controversy in Twitter using a query-based approach for data retrieval and proposes a prediction model which estimates the possibility for a topic to raise controversy in the future. We consider the problem of controversy prediction as a binary classification problem, and propose a logistic regression model to predict whether a topic is to become controversial or not. After pre-processing the collected tweets, they are classified in the context of sentiment analysis. Next, a variety of features expressing different characteristics of the tweets, such as linguistic and temporal information, are extracted for the purposes of our work. We propose aggregating sets of tweets, instead of considering each tweet separately, and extracting aggregated features that are semantically richer. Using logistic regression the statistically significant features are selected and used for the classification. Our experimental results show that the model can achieve 77% accuracy and that statistically significant features express different characteristics strengthening our approach.