Link Prediction in Academic Cooperation Networks

科研合作網絡中的鏈接預測研究

Student thesis: Doctoral Thesis

View graph of relations

Author(s)

Related Research Unit(s)

Detail(s)

Awarding Institution
Supervisors/Advisors
Award date16 Apr 2018

Abstract

Academic cooperation plays an important role in the current advancement of science and technology. Academic cooperation networks are also becoming popular in promoting academic collaboration and knowledge sharing. Predicting possible collaborations among academics in their research and innovation activities is a challenging research topic because of different classification systems with research projects, publications and patents.

With the use of an academic collaboration network platform (www.ScholarMate.com), this study proposes a social capital theory-based framework with structural, relational and cognitive network dimensions to analyse features for possible collaborations among academics. The academic community holds important topological information. Thus, neighbour-based link prediction features are extended on the basis of the global network structural dimension. Institutional connectivity and trust relationships are mined for the relational dimension. The unigram language model and author-latent Dirichlet allocation model are used for the cognitive dimension to accurately match the study interests of researchers.

A new algorithm called four-variable sequential minimal optimisation (FV–SMO) is introduced for large-scale social network mining. Based on the support vector machine (SVM) classification model, this model is derived by solving a series of QP problems with four variables at each iteration via maximal violating pair to ensure that FV–SMO quickly approaches the optimal solution. A theorem is introduced to SVM training to guarantee analytical solutions to corresponding sub-problems.

The proposed framework and algorithms are used for link prediction in an academic cooperation network constructed using experimental data collected from Web of Science. Two link prediction datasets are constructed using two sampling strategies, namely, one to one and one to fifty. The experiment is divided into three stages. In the first stage, the efficiency of FV–SMO is evaluated; the experimental results show the superiority of FV–SMO on the real cooperation network. In the second stage, link prediction performance in the three dimensions is evaluated. In the third stage, all collaborative features of the three dimensions are integrated and different classification models are constructed for link prediction in the academic cooperation network. Results demonstrate the rationality of the constructed cooperation features and the high accuracy of the link prediction model.

    Research areas

  • Academic cooperation networks, SVM, FV-SMO, Language model