Identifying Topic-Specific Opinion Leaders Through Graph Embedding

Abstract

In social networks, some users have an exceptional ability to affect the opinions and behaviours of others. Such people are known as opinion leaders. Accurately identifying opinion leaders can assist greatly in the study of information flow throughout social networks, in addition to providing valuable insights for marketing purposes. Furthermore, contemporary graph embedding tools can significantly improve the process of identifying opinion leaders. Despite this, little research has focused on combining graph embedding and opinion leader detection. Consequently, this thesis focuses on research that integrates these two areas together. I do this by first creating graph embeddings that capture the information contained in the data. Then, I feed these embeddings to an opinion leader detection algorithm that is designed to use the information captured by the embeddings to create a ranking of users, with the highest-ranking users being the designated opinion leaders. I compare these results with several benchmarks of opinion leader detection methods that do not use graph embeddings. The quality of opinion leaders is similar for each method, though not all models designate the same users as opinion leaders. To conclude, I discuss this research in the broader context of the field of graph embedding and opinion leader detection, and provide suggestions for further research.