Article Preview
Top1. Introduction
Recommendation systems (RS) have proven their ability to solve the information overload problem by retrieving items that the user might find interesting (Ricci et al., 2011).
Several techniques and approaches exist in the literature to build a RS. The content-based approach has been extensively used for providing personalized suggestions to the user. The general principal of this approach consists of analyzing the content of items previously liked by the user, and, by means of a similarity measure, recommending items that are the most similar to the user preferences (Adomavicius & Tuzhilin, 2005).
However, despite the advances in RSs some problems still challenging this technology. One of the known challenges is directly related to the fact that traditional RSs are generally designed to only work with data (items/users) that are locally stored in the RS database. This kind of design raises problems associated to the lack of rating information (i.e. cold-start), especially for new users. Indeed, when a user newly registers, the system has no knowledge about his tastes, which may degrade the overall recommendations quality.
To overcome this problem, one promising solution was to gather information from outside the system (Berkovsky, Kuflik, & Ricci, 2008; Cantador, Fernández-Tobías, Berkovsky, & Cremonesi, 2015). In this sense, Online Social Networks (OSNs) have been used as an auxiliary data source for RS, such as in (Abel, Herder, Houben, & Henze, 2013), since users are increasingly using different OSNs in which they permanently leave footprints (e.g. social networks information). Such information can be extremely useful to make recommendations in a third-party application, addressing the cold-start problem.
However, exploiting data from auxiliary sources raises the issue of heterogeneity in the nature of data between the auxiliary and the target domains (Cantador et al., 2015). For instance, data (items/users) available on Youtube platform are not necessarily the same on other OSN platforms. Therefore, the solution for this kind of cross-domain RS was to introduce semantics. In this regard, background knowledge from Linked Open Data (LOD) datasets (e.g. DBpedia) has been explored in literature and gave good results (Orlandi, 2012; Piao & Breslin, 2016a). In fact, thanks to the domain agnostic nature of LOD datasets, such as DBpedia, user interests can be represented with a LOD-resource based model and next be transported from auxiliary to target OSN. In this regards, LOD-based similarity measures are needed to compare resources, especially resources of different domains (i.e. without a domain restriction).
Different LOD-based similarity measures have been proposed in the literature, varying on the type of LOD information they exploited. A first kind of the state-of-the-art measures, such as (Passant, 2010; Piao, Ara, & Breslin, 2015), relied mainly on the concept of connectivity to measure the degree of relatedness between resources, while a second kind of measures, such as (Cheniki, Belkhir, Sam, & Messai, 2016), measured the similarity of resources by comparing their description (i.e. properties, classes, categories). However, relying only on the concept of relatedness might not always reflect the true similarity between resources, while relying only on the similarity of descriptions might fails, due to the lack of completeness in the description of some LOD resources (Zaveri, Maurino, & Equille, 2014).
In this article, the authors focus on investigating the similarities between LOD resources, and make the following contributions: