A Large Multilingual and Multi-domain Dataset for Recommender Systems

A Large Multilingual and Multi-domain Dataset for Recommender Systems Giorgia Di Tommaso, Stefano Faralli*, Paola Velardi Department of Computer Science,*Unitelma Sapienza, University of Rome, Italy {ditommaso,velardi}@di.uniroma1.it, stefano.faralli@unitelmasapienza.it Abstract This paper presents a multi-domain interests dataset to train and test Recommender Systems, and the methodology to create the dataset from Twitter messages in English and Italian. The English dataset includes an average of 90 preferences per user on music, books, movies, celebrities, sport, politics and much more, for about half million users. Preferences are either extracted from messages of users who use Spotify, Goodreads and other similar content sharing platforms, or induced from their ”topical” friends, i.e., followees representing an interest rather than a social relation between peers. In addition, preferred items are matched with Wikipedia articles describing them. This unique feature of our dataset provides a mean to derive a semantic categorization of the preferred items, exploiting available semantic resources linked to Wikipedia such as the Wikipedia Category Graph, DBpedia, BabelNet and others. Keywords: social mining, recommender systems, Twitter, users’ interest dataset 1. Introduction Recommender systems are widely integrated in online ser- vices to provide suggestions and personalize the on-line store for each customer. Recommenders identify preferred items for individual users based on their past behaviors or on other similar users. Popular examples are Amazon (Linden et al., 2003) and Youtube (Davidson et al., 2010). Other sites that incorporate recommendation engines in- clude Facebook, Netflix, Goodreads, Pandora and many others. Despite the vast amount of proposed algorithms, the evalu- ation of recommender systems is very difficult (Fouss and Saerens, 2008). In particular, if the system is not opera- tional and no real users are available, the quality of recom- mendations must be evaluated on existing datasets, whose number is limited and what is more, they are focused on specific domains (i.e, music, movies, etc.). Since differ- ent algorithms may be better or worse depending on the specific purpose of the recommender, the availability of multi-domain datasets could be greatly beneficial. Unfor- tunately, real-life cross-domain datasets are quite scarce, mostly gathered by ”big players” such as Amazon and eBay, and they not available to the research community1. In this paper we present a methodology for extracting from Twitter a large dataset of user preferences in multiple do- mains and in two languages, Italian and English. To reli- ably extract preferences from users’ messages, we exploit popular services such as Spotify, Goodreads and others. Furthermore, we infer many other preferences from users’ friendship lists, identifying those followees representing an interest rather than a peer friendship relation. In this way we learn, for any user, several interests concerning books, movies, music, actors, politics, sport, etc. The other unique feature of our dataset, in addition to multiple languages and domains, is that preferred items are matched with corre- sponding Wikipedia pages, thus providing the possibility to generalize users’ interests exploiting available semantic 1https://recsys.acm.org/wp-content/uploads/2014/10/ recsys2014-tutorial-cross_domain.pdf resources linked to Wikipedia, such as the Wikipedia Cate- gory Graph, Babelnet, DBpedia, and others. The paper is organized as …

Content truncated to 500 words