이기종 그래프 기반 링크 예측 모델을 통한 블록 코딩 추천

In this study, we propose a graph-based link prediction model for recommending meaningful information to users in a coding education platform. To effectively utilize the connection relationships between users, projects, and blocks with different characteristics, we model them using a heterogeneous graph structure. To demonstrate the effectiveness of the proposed heterogeneous graph-based link prediction model, two problems are defined: 1) project topic recommendation and 2) code block recommendation. First, to recommend project topics preferred by users, we cluster individual projects into project topic nodes. This is achieved by clustering projects into identified topics using the Latent Dirichlet Allocation algorithm. Therefore, the defined graph model consists of two types of nodes: 1) user nodes and 2) project topic nodes. User nodes have features such as user type, written posts, and text information from posts of other users preferred by the user. Next, to recommend blocks to users, a new heterogeneous graph is modeled with each defined as a different type of node. In this case, user nodes have the same features as the graph for project topic recommendation. Ultimately, both graph models recommend project topics or blocks that users would prefer based on the information of each data and the connection relationships between different types of data. Through experiments, we confirm that the proposed heterogeneous graph-based link prediction model for project topic recommendation achieves a maximum performance improvement of approximately 58.2% and 85.3% compared to neural network-based classification models such as DNN and RNN, as measured by the AUC score, respectively.