Metapath and attribute-based academic collaborator recommendation in heterogeneous academic networks

Academic collaboration is fundamental to the advancement of scientific research. However, with the growing number of publications and researchers, it becomes increasingly challenging to identify suitable collaborators. Academic collaborator recommendation is a promising solution to this problem. Traditional recommendation methods based on collaborative filtering suffer serious data sparsity. In recent years, network topology-based methods have shown good recommendation performance while alleviating the data sparsity issue to some extent by exploiting the relationships between nodes and their attributes. Nevertheless, these methods are typically based on homogeneous collaboration networks that consist only of scholar nodes and collaboration relationships, leading to suboptimal performance. In reality, collaboration involves many different types of nodes and relations that accumulate multiplex information. To address this issue, we construct a heterogeneous academic information network comprising four types of nodes: scholars, papers, organizations, and publication venues. An academic collaborator recommendation model is designed to capture multi-type attribute features and network topology features of nodes through metapaths based on the network. Specifically, the attribute features of nodes are embedded by a node type-aware embedding method. The topology features are then extracted through the node type-aware aggregation and metapath instance aggregation procedure. After that, we utilize a metapath aggregation method to gather different types of metapaths, each representing a factor that affects collaboration. Thus, the topology information and attribute information are preserved, while encompassing multi-type factors of collaboration. Finally, we compute the vector similarity to determine collaborators. Through rigorous experimentation on a large-scale interdisciplinary academic dataset, we found that the proposed model exhibits outstanding performance in practical applications. Unlike traditional approaches confined to homogeneous collaboration networks, our model delves deeper by mining and leveraging diverse node attributes and multiple collaboration influencing factors. This approach significantly enhances the accuracy and effectiveness of collaborator recommendations. Ultimately, we aspire to contribute to a more efficient and accessible platform that simplifies the search for suitable collaborators.