Social networks enable us to stay in contact with people all over the world, and to read news related to events that occur over the globe in almost real-time. While this has many positive impacts, such as allowing the report of adverse events. Unfortunately, such speed of diffusion of information over social networks has also important negative implications. In many cases malicious users may try to take advantage of the structure of social networks to spread misinformation. Such news can be really convincing that even user without bad intents may end-up sharing and spreading misinformation. To contrast such aspect, it is of crucial importance to understand how misinformation spreads over social networks, and in particular identifying all those nodes that systematically spread misinformation over social networks.

A new Challenge. In the current big data era, content over social networks is produced at an unprecedent rate. While companies that rule such complex systems put significant effort in contrasting the spreading of misinformation, many users get exposed to false or misleading information. This has a practical impact in the current society, since misinformation can be used to bias or polarize opinions on many controversial themes such COVID-19 or political elections, that have practical impacts on everyday lives. Many approaches have been proposed to contrast the spread of misinformation, such as for example text-based approaches that aim at flagging misinformation based on text processing. While this and other approaches can filter some misinformative content, some text cannot be filtered since it appears fact-checked, e.g., contents may intentionally distort reality even quoting scientific references. Therefore, novel techniques that consider the underlying patterns of the spreading process carried by the users, to complement the existing approaches are needed.

Social networks as temporal networks and patterns. Social networks can be modeled as graphs or networks and many algorithmic approaches can be used on such model to infer desired properties of the network. Unfortunately, usual (static) networks do not account for the timing of occurrence of the events in a network. The timings of occurrence of event in fact are related to how information spreads over a network, for example if three users are connected by a path (i.e., there is a flow of information spreading from the first to the last user through the middle user) but the timing of the first edge on the path is greater than the timing on the second edge on the path then information cannot flow on such path. Accounting for timing of events can provide us a new lens for identifying important properties of the social networks modeled through temporal networks. Temporal patterns are defined as frequent structures that occur repeatedly in a short amount of time over a temporal network, these patterns capture both the way users interact over the network and the dynamics through which such interactions are performed. By analyzing different patterns, we are able to understand if a user behavior is frequent or infrequent and how it compares to the different patterns tested. Therefore, through such analysis we are able to correlate users with important functions over the social network.

Temporal patterns and algorithms to tackle misinformation. In my experience at KTH, working with Professor Aristides Gionis, I had the possibility to model the problem of detecting how misinformation spreads across social networks with temporal networks and temporal patters. Since a temporal pattern can encode several ways of spreading misinformation over a social network, we worked on the problem of identifying those users that most contribute to realizing such patterns. This is a novel problem, never tackled in literature on temporal networks and it has several challenges, including scaling the computation on billion edges networks, and providing rigorous theoretical guarantees on the quality of the solution computed. We were able to address this problem and to provide new algorithmics tool to identify nodes that contribute to a high percentage of actions that lead to the spreading of misinformation across social networks modeled as temporal networks. We provide rigorous algorithms to address this problem and we believe that such algorithms will have a practical impact in many scenarios, since we also addressed the problems of scalability and efficiency mentioned previously.

Summarizing. Thanks to the SoBigData++ TNA fellowship I had the possibility to visit the research group of Professor Aristides Gionis and carry on a research project in a very stimulating environment. At KTH I met many researchers working on topics related to my research interests and had the possibility to share and discuss many ideas. Working with Prof. Gionis we developed several algorithms to detect the spread of misinformation by leveraging on temporal patterns, we also think that such algorithms will be of practical impact in many scenarios given that such problem was never addressed given its complexity.

This work has been done during a TNA expericence at KTH Royal Institute of Technology in Stockholm.

Author: Ilie Sarpe