User migration prediction in blockchain socioeconomic networks using graph neural networks

The growing popularity of online social media (OSM) has led to the creation of a wide amount of social media platforms. In this context, the increasing competition among platforms and the emergence of decentralized alternatives such as Blockchain Online Social Media (BOSM), have led to more frequent user migrations: individuals tend to switch platforms in search of improved features, content, or communities. Therefore there has been increasing interest in user migration studies modeling and predicting user migration. However, user migration, especially in blockchain-based platforms remains an understudied problem. Existing methods rely on user activity to derive interaction graphs and then address the user migration prediction problem as a node classification task, where user decisions are encoded as node labels. While the performance look promising, there are currently two important research gaps: i) there is no work using graph neural networks, the state-of-the-art in machine learning on graphs; and ii) there is a lack of methods designed to improve prediction performance in the case of class imbalance, i.e. the presence of dominant behavior among the ones to predict. In this paper, we propose a machine learning pipeline utilizing graph neural networks (GNNs) to predict user migration in BOSM. We model the data as a directed temporal multilayer graph, capturing social and monetary interactions among users. To address the problem of class imbalance in node classification, we introduce a data-level balancing technique following an undersampling approach. The evaluation, conducted on data describing user migration across blockchain online social media platforms, shows that graph neural networks are a suitable machine learning approach to perform user migration prediction. Furthermore, the proposed undersampling approach improves predictive power on severely imbalanced data. These results highlight how graph neural networks are effective in predicting user migration, without the need for manual feature engineering and in the absence of user information. Our methodology holds potential for applications beyond user migration, such as fraud detection and bot detection, and opens up venues for further research in other prediction tasks in online social networks and blockchain-based systems.