Supervised and fuzzy rule based link prediction in weighted co-authorship networks

Abstract

Link Prediction is a fundamental problem in the social networks analysis. In order to solve this problem supervised learning algorithms, which include one fuzzy rule based algorithm were applied in this study. Besides supervised learning algorithms, an unsupervised strategy is also applied to compare the supervised and unsupervised results. Two different networks are chosen for the experiments: a computer science co-authorship network and an eye disease co-authorship network. Those networks were used in both weighted and unweighted versions in the experiments. In a network, a weight refers the strength of a relationship between nodes. Our Experiments’ results proved that weighted networks had better results in comparison to unweighted networks. In the link prediction, the task is to predict the new connections in future for unconnected pair of nodes in present. In the link prediction process with supervised algorithms, metric values were employed as predictor attributes and existence of links was used as class labels. On the other hand, in link prediction process with the unsupervised strategy, a ranking method was employed. In the light of the experiments, it was seen that supervised algorithms had better performance than the unsupervised strategy. Furthermore, among the supervised algorithms, a decision tree and a probabilistic algorithm provided the best performances in comparison with fuzzy rule based algorithm.