Tri-Domain pattern preserving sign prediction for signed networks

Abstract

Sign prediction in signed networks is one of the essential parts in signed networks data mining area. However, in most cases, sign information of links in signed networks is insufficient, so transfer learning is necessary to be an efficient tool to help us relieve this challenge. Many methods cannot leverage the knowledge in source domain adequately. Besides, they can hardly overcome the interference of noisy instances or unrelated instances which called negative transfer phenomenon. To decrease the negative transfer phenomenon in existing methods, a sign prediction model by Tri-Domain Relationship Pattern (SP-TDRP) is proposed. TDRP selects an intermediate domain to be the bridge to transfer knowledge from source domains to the target domain. Then TDRP selects the interference instances and eliminates them to make the transferable knowledge more complete and purer. Thus, the sign classifier can be trained via transferable knowledge and predict the signs in the target domain. The considerable efficiency of SP-TDRP is evaluated on some realistic signed networks when compared with other state-of-the-art models.