Computing the semantic similarity of resources in dbpedia for recommendation purposes

Abstract

The Linked Open Data cloud has been increasing in popularity, with DBpedia as a first-class citizen in this cloud that has been widely adopted across many applications. Measuring similarity between resources and identifying their relatedness could be used for various applications such as item-based recommender systems. To this end, several similarity measures such as LDSD (Linked Data Semantic Distance) were proposed.However, some fundamental axioms for similarity measures such as “equal self-similarity”, “symmetry” or “minimality” are violated, and property similarities have been ignored. Moreover, none of the previous studies have provided a comparative study on other similarity measures. In this paper, we present a similarity measure, called Resim (Resource Similarity), based on top of a revised LDSD similarity measure. Resim aims to calculate the similarity of any resources in DBpedia by taking into account the similarity of the properties of these resources as well as satisfying the fundamental axioms. In addition, we evaluate our similarity measure with two state-of-the-art similarity measures (LDSD and Shakti) in terms of calculating the similarities for general resources (i.e., any resources without a domain restriction) in DBpedia and resources for music artist recommendations. Results show that our similarity measure can resolve some of the limitations of state-of-the-art similarity measures and performs better than them for calculating the similarities between general resources and music artist recommendations.