A Novel Approach for Link Prediction in Academic Collaboration Network: A Third-party Perspective


Student thesis: Doctoral Thesis

View graph of relations


Related Research Unit(s)


Awarding Institution
Award date9 Jun 2020


Academic social network platforms (e.g., Mendeley.com, ResearchGate.com, and ScholarMate.com) have been developed to facilitate research collaboration and knowledge sharing. Duo to the information overload, recommendation systems are implemented to help researchers find suitable collaborators on social network platforms. Existing methods model the collaborator recommendation as the link prediction problem in collaboration social networks. Most of them assume that the features of the paired nodes (mainly proximity) determine the relationship formation. However, according to social contagion theory, the shared neighbors between a pair of potential collaborators (i.e., person directly connected to both collaborators) can also influence the building of the collaboration relationship.

Using social influencing theory, this research proposes a novel link prediction framework from the perspectives of both potential collaborators and their shared neighbors. Similarities between potential collaborators are extracted as paired nodes’ features. Three aspects of shared neighbors are taken into account: similarity with paired nodes, social status in the network, and change of social status after new collaborations. The social status is measured with the betweenness centrality of a node in a collaboration social network.

A wide and deep model in deep learning is proposed to consider both paired features and the influence of shared co-authors for the link prediction in academic social networks. The wide part applies the linear function to mine the paired features. While the deep part constructs a fully connected deep neural network to learn the shared neighbors’ influence. A concatenated layer combines the output features from both the wide and deep parts to obtain the final link predictions.

Two real-world datasets are used to measure the performance of the proposed approach by comparing it with existing popular link prediction methods. To make a comprehensive evaluation, both balanced and unbalanced data sampling strategies are employed. The experiment contains three stages. First, the complexity of the proposed approach is analyzed, and the results show that the proposed approach converges quickly. Second, the influence of shared neighbors on link prediction is evaluated. At last, the proposed approach is compared with other methods, and the results demonstrate the superiority of the proposed approach over the others in terms of f1-score, roc_curve, and auc metrics on the datasets.

This thesis proposes a novel approach to solve the link prediction problem, and it has shown to have enhanced collaborator recommendations in the academic social network platform. From the theoretical perspective, this research considers both the collaboration and competition (i.e., coopetition) aspects of the academic social networks, and it further develops a new link prediction framework with consideration of shared neighbors’ influence. From the practical perspective, the proposed methods can be used for implementation on academic socials network platforms (e.g., ScholarMate) to help scholars find suitable collaborators.