Rumour veracity detection on twitter using particle swarm optimized shallow classifiers

Information overload on Web has been a well-identified challenge which has amplified with the advent of social web. Good, bad, true, false, useful, useless all kinds of information disseminates through the social web platforms. It becomes exceedingly imperative to pro-actively resolve rumours and inhibit them from spreading among the Internet users as it can jeopardize the well-being of the citizens. The task for rumour analysis intends to identify & classify a rumour either as true (factual), false (nonfactual) or unresolved. Determining the accuracy of a rumourous story, a.k.a. rumour veracity is hard owing to the noisy, ambiguous and heterogeneous use of natural language. This necessitates automation of the predictive task which classifies the questionable veracity of rumour accurately. The research presented in this paper, is an empirical study to put forward an optimized learning model which classifies real-time tweets on the basis of truth value, facilitating rumour analysis. The study is conducted on a collection of nearly 14 k tweets pertaining to the recent mob lynching fuelled by rumours on suspected child-lifters in the Indian sub-continent (#moblynching) and run on five classical shallow classifiers to categorize tweets into true, false and unspecified using 13 attributes (features). Subsequently, the use of an optimal feature selection method, particle swarm algorithm is proposed to improve the classifier’s performance. The empirical analysis validates that the proposed implementation of particle swarm optimization (PSO) for feature subset selection in rumour veracity classification outperforms the baseline supervised learning algorithms. An average 11.28% improvement in accuracy and approximately 31% average reduction in features are demonstrated using PSO. The highest accuracy with optimization of 96.15% is achieved by decision tree.