Link Prediction via Multi-hashing Framework

Abstract

Link prediction is crucial in various real world applications such as social network analysis and recommendation systems. For example, in social networks, where social actors and their ties (friendship or collaboration) are represented as nodes and links, link prediction can help anticipate future social tie formation. This problem has generally been tackled through computing a “similarity” – measured through graph topological structure or various node attributes and relationships among them (e.g. researcher’s affiliation or research interest). However, when considering multiple relationships, existing link prediction methods often ignored that similarities across different relationships may be “non-transitive”, i.e., they are not necessarily consistent with each other. Here, we develop a semi-supervised link prediction method via a Multi-Component Hashing framework. We derive multiple hashing tables for nodes in a network with each hash table corresponding to a particular type of non-transitive similarity aspect such as prior collaboration experience or topical interest. New links are predicted based on whether nodes are closer in the hashing tables. Results on three co-authorship networks show that our approach outperforms the state-of-the-art unsupervised and supervised methods. The results also show the superiority of our method in cold-start link prediction setting, where no or little knowledge about the nodes’ network positions is given in the training phase.