Enhancing link-based similarity through the use of non-numerical labels and prior information

Several key applications like recommender systems require to compute similarities between the nodes (objects or entities) of a bipartite network. These similarities serve many important purposes, such as finding users sharing common interests or items with similar characteristics, as well as the automated recommendation and categorization of items. While a broad range of methods have been proposed to compute similarities in networks, such methods have two limitations: (1) they require the link values to be in the form of numerical weights representing the strength of the corresponding relation, and (2) they do not take into account prior information on the similarities. This paper presents a novel approach, based on the SimRank algorithm, to compute similarities between the nodes of a bipartite network. Unlike current methods, this approach allows one to model the agreement between link values using any desired function, and provides a simple way to integrate prior information on the similarity values directly in the computations. To evaluate its usefulness, we test this approach on the problem of predicting the ratings of users for movies and jokes.