With the growth of social media and user-generated content in the recent decade, people are more exposed to the news with unknown sources and fake content. However, validating the credibility of such information is not a trivial task for the majority of the users, and spreading false information could potentially lead to losses and crimes. As a result, it is essential to develop accurate techniques for distinguishing between fake and real news. Our contribution is proposing a multi-modal approach that aggregates the hidden representation of textual news using a variational autoencoder and topic-related features inferred from Latent Dirichlet Allocation (LDA) mixture model to achieve a more accurate and interpretable model. Due to the absence of multimedia and information about the author and spread pattern in many real-world news sources, we focus on extracting relevant features only from the textual content.
