Academic collaborations: a recommender framework spanning research interests and network topology

Fruitful academic collaborations have become increasingly more important for solving scientific problems, participating in research projects, and improving productivity. As such, frameworks for recommending suitable collaborators are attracting extensive attention from scholars. In an effort to improve on the current solutions, we have developed an approach that produces recommendations with better precision, recall, and accuracy. Our strategy is to comprehensively consider the similarity of both scholars' research interests and their collaboration network topologies, leveraging the benefits of these two common similarity indicators into one unified collaborator recommendation framework. A Word2Vec model creates word embeddings of research interests, which solves the problem of calculating similarity solely based on co-occurrence, not context, while a Node2Vec model automatically extracts and learns the topological features of a co-authorship network, moving beyond just local features to capture global network features as well. Then the CombMNZ method is used to fuse the results of the two similarity measures. A ranked collaborator list is then generated to recommend potential collaborators to the target scholars. The workings of the framework and its benefits are demonstrated through a case study on academics in the field of intelligent driving and a comparison with the three baselines: Random Walk with Restart (RWR), Latent Dirichlet Allocation (LDA), and Researcher’s Interest Variation with Time (RIVT). Our framework should be of benefit to academics, research centers, and private-enterprise R&D managers who are seeking partners. We hope that, through the framework’s recommendations, collaborators will form strong partnerships and be able to achieve the ultimate goal of completing research projects, solving scientific problems, and promoting discipline development and progress.