其他摘要：In this study, we propose a researcher recommendation model that considers the roles of principal researchers and co-researchers in research collaborations using a directed researcher network. The proposed model defines edge vectors between researchers, constructs the directed researcher network using the grants-in-aid for scientific research database, and outputs recommendation lists of lists of top-k researchers according to recommendation scores calculated by logistic regression. The recommendation accuracy of normalized Discounted Cumulative Gain (nDCG) @k for the proposed five baseline models are compared. Adamic/Adar, Hasan model, CCRec, Araki model, and LDAcosin are used as baselines. The experimental result shows that nDCG@k, recommendation accuracy, improves from the baseline models for recommendation of principal researchers who have belonged to the same institution. In the other models except CCRec, the nDCG@k by co-researchers recommendation is higher than that principal researchers recommendation; therefore, the difficulty of prediction is considered to be different. It suggests that the co-researchers recommendation may be easily predicted by considering directions of edges. The proposed model recommends researchers with a high average PageRank for the principal researchers recommendation. It also recommends researchers with a large number of collaborative researches. Due to the difference in the recommendation list between the proposed model and the Hasan model, it is possible to recommend a combination of multiple models when implementing as a recommender system. In the logistic regression analysis of features, it is confirmed that there is a difference in the odds ratio between the number of common principal researchers and the number of common co-researchers. These features can only be defined in directed graphs. It is indicated that the proposed model takes into account the differences between principal researchers and co-researchers.
