Yu Qi, Long Chao, Lv Yanhua, Shao Hongfang, He Peifeng, Duan Zhiguang
Department of Medical Information Management, Shanxi Medical University, Taiyuan, Shanxi, China; School of Public Health, Shanxi Medical University, Taiyuan, Shanxi, China.
School of Medicine, Stanford University, Stanford, California, United States of America.
PLoS One. 2014 Jul 3;9(7):e101214. doi: 10.1371/journal.pone.0101214. eCollection 2014.
Research collaborations are encouraged because a synergistic effect yielding good results often appears. However, creating and organizing a strong research group is a difficult task. One of the greatest concerns of an individual researcher is locating potential collaborators whose expertise complement his best. In this paper, we propose a method that makes link predictions in co-authorship networks, where topological features between authors such as Adamic/Adar, Common Neighbors, Jaccard's Coefficient, Preferential Attachment, Katzβ, and PropFlow may be good indicators of their future collaborations. Firstly, these topological features were systematically extracted from the network. Then, supervised models were used to learn the best weights associated with different topological features in deciding co-author relationships. Finally, we tested our models on the co-authorship networks in the research field of Coronary Artery Disease and obtained encouraging accuracy (the precision, recall, F1 score and AUC were, respectively, 0.696, 0.677, 0.671 and 0.742 for Logistic Regression, and respectively, 0.697, 0.678, 0.671 and 0.743 for SVM). This suggests that our models could be used to build and manage strong research groups.
鼓励开展研究合作，因为往往会产生协同效应并取得良好成果。然而，创建和组织一个强大的研究团队是一项艰巨的任务。个体研究人员最关心的问题之一是找到专业知识能与自己形成最佳互补的潜在合作者。在本文中，我们提出了一种在共同作者网络中进行链接预测的方法，其中作者之间的拓扑特征，如阿当米克/阿达指数（Adamic/Adar）、共同邻居、杰卡德系数（Jaccard's Coefficient）、优先连接、卡茨β指数（Katzβ）和传播流（PropFlow），可能是他们未来合作的良好指标。首先，从网络中系统地提取这些拓扑特征。然后，使用监督模型来学习在确定共同作者关系时与不同拓扑特征相关的最佳权重。最后，我们在冠状动脉疾病研究领域的共同作者网络上测试了我们的模型，并获得了令人鼓舞的准确率（逻辑回归的精确率、召回率、F1分数和AUC分别为0.696、0.677、0.671和0.742，支持向量机的分别为0.697、0.678、0.671和0.743）。这表明我们的模型可用于建立和管理强大的研究团队。
