Mining semantic information of co-word network to improve link prediction performance

Link prediction in co-word network is a quantitative method widely used to predict the research trends and direction of disciplines. It has aroused extensive attention from academia and the industry domain. Most of the methods to date predicting co-word network links are only based on the topology of the co-word network but ignore the characteristics of network nodes. This paper proposes an approach with an attempt to exploit network nodes’ semantic information to improve link prediction in co-word network. Our work involves three major tasks. First, a new semantic feature of network nodes (based on the original technology) was proposed. Second, multiple ground-truth data sets which consist of literature from the Information Science and Library Science, Blockchain and Primary Health Care fields are built. Third, to validate the effectiveness of the new feature and prior ones, extensive prediction experiments are carried out based on the data set we construct. The result shows that the new predictive models with semantic information obtain more than 80% of overall accuracy and more than 0.7 of Area Under Curve, which indicates the effectiveness and stability of the new feature in different feature sets and algorithm sets.