Challenges of computational verification in social multimedia

Fake or misleading multimedia content and its distribution through social networks such as Twitter constitutes an increasingly important and challenging problem, especially in the context of emergencies and critical situations. In this paper, the aim is to explore the challenges involved in applying a computational verification framework to automatically classify tweets with unreliable media content as fake or real. We created a data corpus of tweets around big events focusing on the ones linking to images (fake or real) of which the reliability could be verified by independent online sources. Extracting content and user features for each tweet, we explored the fake prediction accuracy performance using each set of features separately and in combination. We considered three approaches for evaluating the performance of the classifier, ranging from the use of standard cross-validation, to independent groups of tweets and to cross-event training. The obtained results included a 81% for tweet features and 75% for user ones in the case of cross-validation. When using different events for training and testing, the accuracy is much lower (up to %58) demonstrating that the generalization of the predictor is a very challenging issue.