Similarity-based Common Neighbor and Sign Influence Model for Link Prediction in Signed Social Networks

Numerous advances have been made in prediction for signed networks. However, few methods can perform simultaneous link and sign prediction. Particularly, for networks with special topologies, the prediction performance of negative links is poor. Moreover, methods based on common neighbors rely heavily on the local structure to achieve a high prediction accuracy. Therefore, a novel model was proposed to realize link and sign prediction and improve the prediction accuracy of negative links. First, the concept of sign influence based on structurally balanced rings was introduced. Subsequently, the similarity of the two nodes based on their first-and second-order common neighbors was defined. Finally, the method was optimized by adjusting the step-size parameters. Experiments were performed on six classic datasets. The prediction accuracy of Epinions, Slashdot, and Wikipedia were 95.1%, 92.5%, and 97.7%, respectively. In the network with a 1:1 ratio of positive and negative edges, the sign prediction accuracy can reach 99.3%. The comparison and results of the recommended top k links proved the effectiveness of the proposed algorithm and showed its considerably high link prediction precision and low computational complexity in link and sign prediction for sparse networks and negative link prediction.

Automated summary

A novel model was proposed for simultaneous link and sign prediction, with high accuracy for negative link prediction. Experiments showed the model's effectiveness on various datasets, achieving high prediction accuracy levels.