engleski

A multilayer network approach for fake news detection during the COVID-19 crisis

Digital communication technologies and social media accelerates information spreadig. This trend is reflected in the infodemics – a rapid difusion of information and misinformation. Infodemics is potentially very dangereous, especiallly during the crisis and in emergency situations, because it makes difficult for individuals and groups to find reliable sources of information. COVID-19 outbreak infodemic resulted in massive misinformation and fake news spreading. In that light, automatic detection prediction and preventing fake news spreading plays an important role and may improve various aspects of crisis communication. In last ten years there have been pubblished a vast number of research papers and proposed approaches adrresing the problem of automatic detection of fake news, however it remains a challenging task, especially in the context of COVID-19 crisis communication. The aim of our research is to define an approach that integrate various features of information spreading in social media and thus enables the classification into two categories: true and false. More precisely, we will propose a multilayer framework that defines a set of approaches, methods and network-based models that capture three aspects of information spreading analysis: (i) content, (ii) context and (iii) dynamic. Message content captures semantics and it can be represented as an embedding using deep learning models (such as for example BERT). Message context refer to the features related to the structure/topology of the social network related to the message and can be described true various layers. For example, Twitter messages features extend to reply layer, retweet layer, mention layer and quote layer. To represent all these features in one formalism, we propose multilyer network. The third aspect of messsage is its spreading dynamcs which can be described in terms of how max-breadth, depth, size of diffusion, etc. tend to change during the time. Thus, the main idea of the propsed approach is to extend the message content embedding with other features and to constuct a multilyer embedding which will be used in the classification of messages related to the fake news. Within the proposed framework, we will study empirical data related to COVID-19 crisis communication crawled from various social media sources, such as Twitter or Facebook and online portals. The main focus of our datasets will be texts in the Croatian language, however, to be comparable with other studies, we will perform experiments with texts in the English language as well. We expect that the results of this research will enable a better understanding of the fake news spreading and help in the future development of a systems for detecting and preventing the spread of misleading and harmful information on social media.

fake news detection ; deep learning ; natural language processing ; social networks analysis ; COVID-19

nije evidentirano