Learning multi-resolution representations of research patterns in bibliographic networks

This study aims at representing research patterns of bibliographic entities (e.g., scholars, papers, and venues) with a fixed-length vector. Bibliographic network structures rooted in the entities are incredibly diverse, and this diversity increases in the outstanding entities. Thus, despite their significant volume, the outstanding entities obtain minimal learning opportunities, whereas low-performance entities are over-represented. This study solves the problem by representing the patterns of the entities rather than depicting individual entities in a precise manner. First, we describe structures rooted in the entities using the Weisfeiler–Lehman (WL) relabeling process. Each subgraph generated by the relabeling process provides information on the scholars, kinds of papers they published, standards of venues in which the papers were published, and types of their collaborators. We assume that a subgraph depicts the research patterns of bibliographic entities, such as the preference of a scholar in choosing either a few highly impactful papers or numerous papers of moderate impact. Then, we simplify the subgraphs according to multiple levels of detailedness. Original subgraphs represent the individuality of the entities, and simplified subgraphs represent the entities sharing the same research patterns. In addition, simplified subgraphs balance the learning opportunities of high- and low-performance entities by co-occurring with both types of entities. We embed the subgraphs using the Skip-Gram method. If the results of the embedding represent the research patterns of the entities, the obtained vectors should be able to represent various aspects of the research performance in both the short-term and long-term durations regardless of the performances of the entities. Therefore, we conducted experiments for predicting 23 performance indicators during four time periods for four performance groups (top 1%, 5%, 10%, and all entities) using only the vector representations. The proposed model outperformed the existing network embedding methods in terms of both accuracy and variance.