Recommender Systems (RS) bene t from the richly-structured information contained in publicly available RDF-graphs. Not only is the burden of extracting features from text thereby alleviated, but also the interconnections in the graph have been proven to be useful. These allow one to nd related items in the graph which do not necessarily share a large number of features, e.g. items from di erent domains. Thus, these graphs can be exploited to generate cross-domain recommendations, i.e. to recommend items using feedback provided in a di erent domain. To bene t from RDF's data model, RecSPARQL has been proposed as an extension of SPARQL together with a system which evaluates such queries. Although this solution makes it possible to generate recommendations on top of arbitrary RDF-graphs, it is limited to single-domain recommendations. In this paper we present an extension of RecSPARQL's syntax and semantics for cross-domain recommendations. Our experiments on a very sparse dataset show that the added components can help to improve the quality of recommendations.
Driven by the Semantic Web (SW) initiative, a considerable number of
knowledge graphs are published according to the Resource Description Framework (RDF).
For Recommender Systems (RS) [
Our extension consists of adding mechanisms to specify the source as well as the target domains. By specifying them, the user and item pro les are extended to embed this information. The new multi-domain pro les are the input to the similarity operator which computes, for instance, the neighborhoods, i.e. the most similar users with respect to each user in the dataset. In the experiments, we use the Facebook dataset (2nd LoD-enabled RS Challenge, ESWC 2015), which contains feedback for three domains: movies, books and music (see statistics in the table below). Moreover, items are directly given as DBpedia resources. Using these IRIs, we augmented the dataset with a subgraph extracted from DBpedia by keeping all paths between items up to three hops regardless of the direction.
#Users #Items #Ratings Sparsity (%) Movie Domain 32,159 5,389 638,268 99.631 Books Domain 1,398 2,609 11,600 99.681
Music Domain 52,072 6,372 1,093,851 99.67 Queries. To show the bene ts of having the feedback, features, and interconnections all in a single graph, we gradually introduce more information in each query (Figure 1). Query Q1 is a traditional collaborative- ltering RecSPARQL query based only on liked movies (single-domain). In Q2 (our extension) we introduced preferences from a second source domain, music. Note that the syntax allows one to de ne a domain (?domain rdf:type recsparql:Domain) and to set even multiple source and target domains (recsparql:SourceDomain, and recsparql:TargetDomain, correspondingly). The system adequately processes the additional information by creating the input to the recommendation model and customizing it. In Q3 we additionally introduce with respect to Q2 some common features between movies and music items. To make use of interconnections in the graph, we introduced the processing of property paths in our
Q1 (traditional RecSPARQL query): 1 RECOMMEND ?user ?item.REC (AVG(?RATING) 2 AS ?avgRat) USING CF 3 WHERE f 4 ?user ex:likes ?item . 5 ?item rdf:type ?type . 6 FILTER(?SIMscore > 0.0 && (?type = dbo:Film jj 7 ?type = sorg:Movie jj ?type = rc:Movie CW) g 8 BASED ON f 9 ?user ex:likes ?item g 10 SETTINGS f (*) g 11 GROUP BY ?user ?item.REC
Q2 (our extension): 1 RECOMMEND ?user ?item.REC (AVG(?RATING) 2 AS ?avgRat) USING CF 3 WHERE f 4 ?user ex:likes ?item . 5 ?item rdf:type ?domain . 6 FILTER(?SIMscore > 0.0) g 7 BASED ON f 8 ?user ex:likes ?item g 9 SETTINGS f (*), (.), 10 dbo:Film rdf:type recsparql:SourceDomain , 11 sorg:Movie rdf:type recsparql:SourceDomain , 12 rc:Movie CW rdf:type recsparql:SourceDomain , 13 rc:MusicalPerformer rdf:type recsparql:SourceDomain , 14 rc:Band MusicGroup rdf:type recsparql:SourceDomain , 15 dbo:Band rdf:type recsparql:SourceDomain , 16 sorg:MusicGroup rdf:type recsparql:SourceDomain , 17 dbo:Film rdf:type recsparql:TargetDomain , 18 sorg:Movie rdf:type recsparql:TargetDomain , 19 rc:Movie CW rdf:type recsparql:TargetDomain g 20 GROUP BY ?user ?item.REC cross-domain extension. This allows one to conveniently specify in the Based On clause any possible path between items from di erent domains. Our extensions uses the additional information in Q2 and Q3 to change the way user pro les are built. Consequently, the system directly alters the recommendation model based on the new neighborhoods. To choose the feature types used in Q3 (genre, label, and subject) we prioritized relevant ones, which were common to both domains and also show a high degree of interconnectivity. The relevance is based on Principal Component Analysis (PCA) and Information Gain (IG). These feature selection techniques measure how representative a feature type is for its corresponding domain. The interconnectivity is based on counting the number of item pairs from di erent domains which are connected by a feature of the same type. All queries were evaluated in a distributed fashion using Spark to alleviate the scalability challenge related to the neighborhood computation. Results. Figure 2(A-B) shows the comparative results with respect to two metrics, mean reciprocal rank (MRR) and normalized discounted cumulative gain (NDCG). MRR tells us how well the recommendation model is doing in nding the
rst relevant item in the list of recommendations. NDCG gives a score for the complete list instead, taking into account the number of hits with respect to the test set but also with respect to the position of the hits in the list. 6 e4 r o c S2 0 5 10 SD-Q1 2 1 0 6 s r 4 o b h g 2 i e N 0 # 104 (C)
104 (D)
10 2(E)
The plots clearly show, for di erent number of recommendations k, the benets of introducing the additional graph components in each query. For instance, when we focus on a single metric, we found that from Q1 to Q2, the number of users with NDCG@50 equal to zero was reduced by 3.79%. From Q2 to Q3 this was reduced again by 3.3%, i.e. overall the reduction was 7.09%.
This was due to changes in the characteristics of the neighborhoods, which can be better understood by looking at the plots (C-E) in Figure 2. These plots show that the average number of neighbors in each neighborhood increases the more information is utilized to build them (C). Moreover, the standard deviation (D) shows that introducing a second domain a ects only some users (probably those with common ratings in both domains). This inequality is later corrected when the features are included. Overall, the average similarity scores between the neighbors and active user consistently decreases, as the plot (E) shows. 25 CD-Q2
We believe there is still room for new tools on top of RDF-graphs to enable useful applications. While leveraging both the graph's content and structure continues to be a challenge, we showed that our extension of RecSPARQL for cross-domain recommendations ful lls its purpose of enhancing the quality of recommendations. Not only did feedback from the additional information source have a positive impact, but the features speci ed by means of property paths also helped to build fairer neighborhoods for the users.