Evidential link prediction by exploiting the applicability of similarity indexes to nodes

Owning to its extensive range of applications, the study of link prediction has captured considerable attention from researchers. Recently, hybrid similarity methods, which incorporate multiple sources of information, have been reported to perform link prediction. However, the applicability of a similarity method to different nodes has not been exploited in the previous hybrid methods. In this regard, we propose a new hybrid algorithm that fuses the results of multiple similarity indexes via the evidence theory. In the proposed algorithm, each similarity index is considered as a source of evidence and the corresponding basic belief assignment (BBA) is derived from the similarity score. More importantly, to adaptively estimate the reliability of evidence, the applicable coefficient (AC) of an index to a node is computed. Then, the BBA of a node pair with respect to a similarity index is discounted according to the ACs of the index to two endpoints. The connection probability of the node pair is gauged by fusing multiple discounted BBAs. Experimental results based on several real-world networks suggest that the proposed method is superior to individual similarity indexes and baseline methods.