<![CDATA[Detecting influencers in a social network of massive student activities is vital for universities because it will help them understand potential leaders and social behavior. This paper mitigates the issues of classical topology-based metrics by presenting volume calculation through Graph Attention Networks (GATs) applied to a real network with 2,520 students and about 282,000 interactions. A new hybrid method of influencer ranking proposed, which combines the node embeddings obtained by GAT with a structural influence signal from PageRank. The evaluation system includes two main parts. First, qualitative evaluation of the hybrid ranking method against PageRank-only. This assessment learns from a ground truth dataset of 993 formal leaders. Second, evaluate the communities found by GNNs against those discovered by classical methods using internal quality criteria, including modularity and conductance. From the observation, PageRank baseline does slightly better than the hybrid method in ranking and both methods are significantly better from a random rank with their Spearman’s Rank Correlation equal to 0.513 for PageRank based and 0.451 of the hybrid variant, respectively. Yet, in the task of community detection, GNNs have greater representational capacity. Even though the resulting modularity score was also very competitive, communities had much lower (and hence better) average conductance than Louvain and Walktrap methods (0.137 vs 0.198 and 0.302). These paired results shows that: the success of a PageRank baseline is tied to our formal-role-based ground truth which is structural. The GNN’s increased ability to discriminate such well-delineated, socially close communities implies that the embeddings it learns better represent the network’s true social structure. In conclusion, while PageRank effectively reveals the formal leaders in a community, our hybrid GAT technique acts as complement to shed light on emerging influencers.]]>
