Applying graph neural networks for stance detection over online social networks

Abstract

Fake news has flooded social networks and has been studied from different perspectives. Different methodologies have been proposed to approach this issue. In recent years, researchers have been conducting analyses based on machine learning and neural network models. The results of a model depend not only on the structure of the model by itself, but also on the right pre-processing of the data. For this reason, various techniques have been developed. The concern of researchers is to achieve the best results through the combination of the appropriate processing of the data with the appropriate model. The current MSc thesis examines the techniques developed in the field of assessing news based on rumors in social networks; to this end, the thesis implements an approach based on the Graph Neural Network algorithm (GNN). In the introductory part of the thesis, the techniques that have already been proposed by researchers are examined. In the next section, natural language processing operations are described with reference to embeddings that play a decisive role in transforming data into a format suitable for input into processing models. In particular, the thesis is based on data set analysis through Graph Convolutional Networks (GCN). For this reason, a detailed description of this model is provided; algorithms of the same category as GNNs are also presented. The data set used for analysis is described; this is a data set built specifically for research on the issue of rumor evaluation. The methodology adopted and employed is subsequently described. In particular, the open access data set that has been mined from the social network Twitter, contains rumors that have concerned the public and labels regarding the stance of users towards these rumors. The data set has already been analyzed through various methodologies, yet, not from the perspective of GNNs. The present thesis aims to analyze the same data set using the GNN architecture and comparatively evaluate results with existing, published results, in order to draw conclusions as well as identify limitations and provide feedback for further analysis.