GCN을 활용한 커뮤니티 네트워크의 콘텐츠 인게이지먼트 예측

초록

As interest in community-based marketing increases, vertical commerce platforms-specialized for specific product or consumer groups-are showing remarkable growth in the online commerce market. These platforms extend beyond transactional functions by integrating community-oriented features designed to enhance user engagement and optimize lock-in effects. Active user engagement within these communities leads to strong loyalty, increasing the platform’s economic value, making it a crucial factor in platform management strategies. However, accurately predicting which particular content within extensive and intricate networked communities will most effectively enhance user engagement remains a considerable challenge. Therefore, this study proposes an approach utilizing Graph Convolutional Networks (GCN) to effectively predict engagement on vertical commerce platforms. The study seeks to examine whether incorporating structural information of communities, which captures complex interactions, enhances the accuracy of engagement prediction. Additionally, the study aims to inform the development of operational strategies that optimize user engagement on the platform. Based on data from “Today’s House”, a representative vertical commerce platform, we conducted quantitative predictions of content engagement metrics (likes, views, comments, saves, shares). We constructed a heterogeneous graph consisting of various node types (users, content, categories) as well as multiple edge types (comments, scraps, uploads) and applied a HeteroGNN model to capture the complex relational structures. Each node was represented by high-dimensional features, including numerical data and SBERT-based text embeddings. Information was aggregated from neighboring nodes and incorporated into the content features to predict content engagement as a multi-target regression problem. The analysis demonstrated that the HeteroGNN model outperformed traditional machine learning models in predicting content engagement. These results empirically demonstrate the importance of relationship-based information in engagement prediction and the structural relationship learning ability of the GCN model, while also suggesting improvements in operational strategies for activating communities within vertical commerce platforms.

키워드