Παραγοντοποίηση Τανυστή με Πληροφορίες Δικτύου για τον Εντοπισμό Ψευδών Ειδήσεων (Network-based Tensor Factorization with Label Information for Fake News Detection)

Abstract

Social media usage for news consumption has grown excessively during the last years. Along with it came the spread of large amounts of "fake news" that are deliberately written to deceive readers. Due to the serious consequences of mass misinformation, fake news detection is considered an urgent technological issue in need of automated solutions. In related literature, we can find several attempts to detect fake news only from its content, but their performance is generally considered non-satisfactory. Hence, we need to incorporate auxiliary information, coming from the network, in order to be able to identify misinformation more accurately. Our intuition is that the network connections of users that share fake news can be discriminatory enough to facilitate the detection of fake news. Consequently, the aim of this thesis is to detect fake news using network-based information and some labeled data. Firstly, we present a novel tensor-based way of representing the news pieces using only the friendships between users who have shared the news. Then, we wish to exploit this representation and a set of labels we have in hand, by developing a tensor factorization method that associates the class information of some posts with their latent representations. Finally, we end up with a unified optimization process that integrates a classification error term inside the factorization itself. Results on real-world datasets demonstrate the effectiveness of our proposed method and suggest that it can become competitive against existing state-of-the-art methods, by employing an arguably simpler approach.