Knowledge Organization Systems for Linked Data ...

3 downloads 1004 Views 212KB Size Report
fitting recommendations with its current discount campaigns. ..... The general workflow of a recommendation process for an input of user preferences is depicted ...
Knowledge Organization Systems for Linked Data-enabled on-the-fly recommendations Lisa Wenige Chair of Business Information Systems Friedrich-Schiller-University of Jena Carl-Zeiss-Str. 3, 07743 Jena

Abstract. Recommender systems (RS) have become an important tool in many e-commerce applications. They help users navigate through rich information spaces. While they often improve user experience and search results, many content-based systems work with insufficient data. Thus, recent research has focused on enhancing item feature information with data from the Linked Open Data (LOD) cloud. The generation of Linked Dataenhanced item descriptions requires a considerable amount of pre-processing which can be a barrier for short update cycles. In addition, item similarity computation is often carried out in offline mode. Thus, Linked Data recommender systems (LDRS) are usually bound to a predefined set of item features and offer limited opportunities to tune the recommendation model on a frequent basis. This is a considerable drawback since knowledge bases on the web offer rich and upto-date information sources, which are well suited for personalization tasks. Above that, they contain statistical information on occurrences of resources, which can be used to measure relevance. This paper introduces the prototype SKOS Recommender (SKOSRec), which produces scalable on-the-fly recommendations through SPARQL-like queries from Linked Data repositories. The SKOSRec query language enables users to obtain constraint-based, aggregation-based and cross-domain recommendations, such that results can be adapted to specific business or customer requirements. Keywords: Linked Data, Recommender Systems, Query-based Recommender Systems

1

Introduction

Since the 1990s recommender systems have become an import application in the area of information filtering. They support users in choosing relevant items and are an efficient tool to deal with the information overload [1].

Among the most known examples of recommender systems are the ones used by wellestablished e-commerce retailers such as Amazon1 or Netflix2. Here, users receive personalized recommendations for items of the product catalogue. Recommender systems usually belong to one of the following types: collaborative filtering, content-based or hybrid systems. Collaborative filtering systems were among the first to be developed [2], [3], [4]. In these applications, users are referred to preferred items of like-minded peers. In content-based systems, recommendations for unknown items are derived by finding similar items that were liked by the user in the past [1]. For similarity computation, a sufficient amount of item feature information has to be processed. An example of an early content-based recommender system can be found in [5]. Both content-based and collaborative filtering systems suffer from certain shortcomings. Whenever a new user is added, there is no data to infer user preferences from (new user problem). The same problem arises when new items are entered into the system (new item problem) or when item metadata is insufficient (limited content analysis). In addition to that, users can receive unfitting recommendations due to sparse preference information (sparsity problem) or information that offers no novelty to the user (overspecialization problem) [1]. The Linked Open Data (LOD) movement gave rise to the type of Linked Data recommender systems (LDRS). These systems tackle drawbacks of traditional approaches by enriching existing recommender systems with information from public data sources, such as DBpedia3. The LOD cloud comprises data on almost any kind of subject and offers general purpose information as well as data from special domains [6]. That is why it is especially well suited for enhancing item metadata in content-based recommender systems to overcome problems of overspecialization and limited content analysis. LOD resources are uniquely identified through URIs. Thus, the LOD cloud provides less ambiguous data sources than text-based item descriptions [7]. Experiments on historic data showed that LDRS are at least competitive with classical systems and sometimes even outperform them in terms of accuracy [7], [8], [9], [10]. But even though current LDRS show promising results, they do not yet take full advantage of the potential that LOD offers. The paper addresses this research gap in the following sections: • Identification of central tasks in typical usage scenarios that cannot be handled by existing RS, but are of high relevance in e-commerce applications (section 2). • Deduction of requirements for an RS that is able to handle the identified tasks by making maximum use of the LOD cloud (section 3). • Survey of current LDRS and query-based RS with a focus on the specificities of the LOD cloud. (section 4).

1

http://www.amazon.de http://www.netflix.com 3 http://wiki.dbpedia.org 2

• Technical description of the SKOSRec prototype, that implements the concepts (section 5). • Summary of the main findings and practical implications of the approach (section 6).

2

Motivation

Existing recommender systems (both LDRS and non-LDRS) usually compute user preferences offline. Thus, the recommendation model is 'hard wired' into the system. It relies on static data and cannot easily be configured. This limited view on the data leads to inflexible recommendation results [11], [12]. For illustration purposes consider the following usage scenarios that were derived from the literature: • Multimedia e-commerce retailer: An e-commerce retailer tracks user preferences through shopping behaviour. While there are many multimedia items (books, movies or music) in the product catalogue, information on their features is sparse in the retailer’s local databases. In addition to that, new items are added every day, which makes manual metadata maintenance costly. Due to insufficient user profile information, the application of common recommendation algorithms does not lead to satisfying results. Above that, the computation of similar users and items does not offer control for adaptation to the retailer’s interest [14]. For instance the retailer aims at fitting recommendations with its current discount campaigns. These campaigns might promote infrequently purchased items of a special type or items that are fit to a certain holiday occasion of the year. • Online travel agency: An online travel agency would like to present recommendations to its customers. But travel activities are usually less frequent than purchases of multimedia items. Additionally, touristic preferences are more complex and cannot as easily be predicted as preferences in other domains [15]. By applying collaborative filtering, two customers would receive similar trip recommendations once they have been to the same cities even though they have visited different points of interest (POI). These two examples illustrate problems of classical recommender systems with the additional requirements of highly personalized recommendations based on fresh data. Central to both scenarios is that they involve a processing step of similarity detection, which is bound to certain restrictions in an update and data intensive environment. As customer and business contexts are subject to constant and frequent changes, it is almost impossible to foresee any preferences that might arise in the future. The above mentioned aspects are currently only partly met by different types of systems, namely: LDRS and query-based RS. Currently, only a few systems are able to produce query-based recommendations (e.g. [13], [12]). For instance, the REQUEST system integrates the personalization process with OLAP-like queries, such that the selection of items/users can be based on certain conditions and aggregations. Thus, recommendation models are customizable to requirements at runtime.

But as information on user preferences is usually sparse, this information becomes even sparser when only certain items or users are selected. This often leads to unreliable recommendation results [13]. In LDRS, the data sparsity problem is tackled by the means of comprehensive item descriptions from the LOD cloud, whereas recommendation models can usually not be configured at runtime. Additionally, the LOD cloud contains billions of triple statements [6], which requires fast processing solutions [17]. Hence, it is aimed at developing a recommender system that integrates the strengths of both approaches and overcomes their weaknesses.

3

System Requirements

Based on the usage scenarios outlined in section 2 and the specificities of LOD, certain requirements (RQ) for a query-based LDRS can be derived (see Table 1 and 2). Opportunity Timeliness Comprehensiveness

Adaptability

Exploratory Search

RQ

Description Many Linked Data repositories offer up-to-date information sources [18], [6] The LOD cloud comprises billions of triple statements ranging from general purpose data to the life sciences domain [6]. Above that, comprehensiveness can be achieved through data integration upon exploring cross-dataset links. Queries over Linked Data sets can be formulated with a high degree of expressiveness, due to the flexibility of RDF [16], [19]. A system that builds upon this will be easily adaptable to changing business or customer requirements. The abundance and heterogeneity of Linked Data can facilitate serendipitous search forms. Thus, navigation along property paths might lead to novel and unexpected search results [20].

Table 1. Opportunities of LOD for recommendation tasks

A A, C, D

C

B, C, D

Problem

Description

Data Quantity

Data Federation

Data Heterogeneity

Due to the large amount of triple information, efficient and scalable strategies for item similarity computation have to be identified. Metadata information might also be spread across several repositories that need to be accessed through different SPARQL endpoints. Retrieval strategies will have to find efficient ways of data access and integration. Numerous vocabularies and data models exist in LOD repositories.

RQ A

B, D

B

Table 2. Problems of LOD for recommendation tasks

The following summary describes the system requirements in detail. • RQ.A: Scalable on-the-fly recommendations. The recommendation model should be adaptable to individual user demands. Recommendations have to be calculated on-the-fly, which requires scalability to a large number of Linked Data resources and user preference information. Thus, similarity values are not to be calculated on the whole item space [17]. Although recommendations should be generated at runtime with low processing times, the system is not designed to produce real-time recommendations. The focus is on the aspect of adaptable recommendation models that can be applied on fresh data with no need for pre-processing. • RQ.B: Flexible and homogeneous similarity detection. The method of similarity calculation should be applicable to heterogeneous data models and different data repositories. A further requirement is that recommendation results can be broadened or narrowed according to different user needs. Hence, result lists can contain either more accurate or more novel and diverse recommendations. • RQ.C: SPARQL Integration. The recommender system should enable efficient navigation on Linked Data knowledge graphs through a combination of similarity calculation and query-like selections. The system has to adhere to principles of the SPARQL 1.1 specification (e.g. graph pattern matching, property paths, subqueries) [21], [22]. Hence, the resource selection process can be based on both precise and imprecise elements. • RQ.D: Cross-repository recommendations. It might often be the case that resources show some similarity, but reside in different Linked Data repositories. The system should be able to map graph patterns between repositories and produce recommendations across datasets.

4

Related Work

The first attempts to use Semantic Web data for recommendation tasks go back to the middle of the last decade [17], [23]. In the following years more and more recommender systems applied Linked Data. There are systems that infer preferences from explicit RDF statements (e. g. [23]), but most LDRS apply a content-based or hybrid approach. In these systems Linked Data is used to complement the information sources at hand. It has been shown, that LDRS can outperform systems that solely rely on local data [9], [24], [7], [8]. But current approaches require a considerable amount of pre-processing. Item features have to be selected and extracted before similarity calculation. Thus, the recommendation model is built offline, which imposes restrictions for ex-post configuration changes. This might also be a reason why most existing recommender systems in research are either from the music or the movie domain (see Table 3). Other areas as well as cross-domain recommender systems are fairly underrepresented [25]. This is unfortunate, since the LOD cloud contains many general purpose datasets that could be used for this kind of task [6]. Table 3 shows that most LDRS and query-based recommender systems are not fully equipped to handle the specificities of the LOD cloud. The following sections outline how the identified requirements are addressed in the SKOSRec prototype.

System [26] [27] [28] [29] [30] [31] [32] [33] [10] [34] [35] [36] [20] [37] [38] [23] [39] [9] [40] [7], [8] [41] [42] [24] [43] [44] [45] [25], [46] [47] [48] [49] [50] [51] [52] [53] [54] [55] [56] [11], [13] [12]

Type RS SPARQL queries RS Exploratory Search RS RS RS RS RS RS RS RS Exploratory Search RS Exploratory Search SPARQL queries RS RS RS RS RS RS RS RS RS RS RS RS RS RS RS RS RS Exploratory Search Query-based RS Query-based RS RS Query-based RS Query-based RS

Domain POI Expert Search Movies, Music Videos Keywords Movies, Stars Movies Museal objects Music Preservation News Web Services General Purpose Events Keywords Music Movies Books Travel Movies Music News Movies, Real Estate Music Music Music POI, Music Citations Movies, Music, Actors Movies Movies Music Movies, Person General Purpose Movies Data Integr., Mapping Scientific Publications General Purpose University Courses

LOD

A

B

( )

( )

( )

Table 3. Survey of LDRS and query-based RS

( )

( )

C

D

5

SKOSRec prototype

5.1

General approach

Most LDRS base the identification of similar items on feature information. But considering the large amount of triple statements in the LOD cloud, using all known features of a resource leads to poor scalability and long processing times. Thus, in the context of LDRS it has been proposed to select certain item features (properties) [7], [8], [37], [48], [28]. But due to the diversity of data models, the selection process can be errorprone and time consuming. Additionally, it requires a profound understanding of the underlying data sets and schema definitions. There are many resource types in repositories such as DBpedia and metadata descriptions differ largely among them. This poses high demands on system designers and hinders wide adoption of Linked Data information for recommender systems as the extracted data might be use-case dependent and overly specific. The problem of data heterogeneity can be addressed by taking advantage of annotations from knowledge organization systems (KOS). A knowledge organization system is a controlled vocabulary (e.g. a taxonomy, classification or subject heading scheme) that helps to organize large collections of items by offering means of disambiguation and hierarchical structures among knowledge concepts. KOS can be described through the Simple Knowledge Organization System vocabulary (SKOS). SKOS provides a standardized schema for organizing conceptual resources in informal hierarchies and association networks. It is based on the Resource Description Framework (RDF) and facilitates exchange, processing and publication of KOS on the Semantic Web. In SKOS vocabularies, knowledge concepts are uniquely identified through URIs and are part of superclass/subclass relationships [57], [58] (skos:borader/skos:narrower). These relationships can be used to broaden/narrow the scope of recommendations. To date many Linked Data sets are annotated with concepts from SKOS vocabularies. They enable handling of data heterogeneity in and across Linked Data repositories through their standardization and expressions for mappings (e.g. skos:exactMatch) [59], [6]. Whereas many LDRS have already applied SKOS annotations [7], [8], [48], these annotations have not been used for query-like on-the-fly recommendations. The following definitions build the foundation of the SKOSRec system. • Definition 1 (KOS annotations) Let be the annotation graph of an RDF dataset ( ⊂ ), where resources are directly linked to concepts of a knowledge organization system via a predefined property (e.g. dct:subject). All nodes of the are IRIs and the annotations of an input resource are defined as follows: (

= {∀ ∈

, ∃< ,

:

, >}

(1)

By retrieving annotations of an input resource , similarity calculation does not have to be performed on the whole item space. Above that, recommendations can be restricted to resources that do match user defined SPARQL graph patterns. The notion of a SPARQL graph pattern is defined in [22].

• Definition 2 (Relevant resources) A mapping Ω of relevant resources and their annotations can be obtained by retrieving all resources that share at least one concept with resource ( ! , , ? # ? $, ,? ). In addition to that, resources are to be excluded when they are not contained in the evaluation of a user defined graph pattern . Ω

%&

|?( )&

∈ * ! +, - ⋈ /&

|?(,?0 1&

∈ 2 345 6

(2)

Upon retrieval of relevant resources and annotations, similarity values can be calculated. They are based on the Information Content (IC) of the shared features of two resources. This idea was introduced by [48], but is expanded to the case when the item space is restricted to match a user defined graph pattern. • Definition 3 (Conditional similarity) Let be the set of KOS features of resource and 7 the set of KOS features of resource 7 and (7 ∈ & ? $)), then their similarity is based on the IC of their shared concepts: 8 ∩ 7 . :;

,7

? 8

(3)

• Definition 4 (Conditional Information Content) The IC of a set of KOS annotations is defined through the sum of the IC of each concept ∈ & ? , where @A 7 is the frequency of and is the maximum frequency among all relevant resources. ? 8

∑0 ∈ L Clog G

H IJ 0 >

K

(4)

The general workflow of a recommendation process for an input of user preferences is depicted in Fig. 1.

Fig. 1. General workflow of the SKOSRec system

5.2

Query Optimization

The extraction of relevant resources and concept annotations can become computationally expensive, especially in cases when concepts are frequently used in a dataset. Hence, it is proposed to reduce the records that have to be retrieved from SPARQL endpoints. By knowing the number of recommendations a user wants to receive, it can be calculated which resources can be omitted without influencing the final ranking. This is the case when the maximum potential score for a certain number of shared features is smaller than the minimum potential score for a higher number of shared features. By this means, it is determined how many annotations have to be shared at least with an input resource (cut value) (see eq. 5 and 6). Ω0MN = /&0MN|?( 1&0MN ∈ O0=M>N Ω

I?M0I?

?0

Ω >

6

= Ω ⋈ Ω0MN

(5) (6)

The effectiveness of the approach was evaluated in the domains of movies, music, books and travel. For this purpose, four datasets were generated each containing 100 randomly selected DBpedia resources. The performance test was carried out on an Intel Core i5 2500, clocked at 3.30 GHz with 8 GB of RAM, where 4 GB of RAM were reserved for a DBpedia 3.9 endpoint running on a local virtuoso server. Table 4 shows that the optimization considerably reduces processing times. Domain Movie Book Music Place

Ø Exec. Time (ms) regular optimized 9961 1921 303 92 4145 501 218 73

regular 23542 1206 7662 621

Ø No. of triples optimized 20535 837 501 134

Table 4. Results of the performance test

5.3

SKOSRec Query Language

The SKOSRec query language uses elements of the SPARQL 1.1 syntax [21] . It enables graph pattern matchings and subquerying. There already exist some language extensions, that combine SPARQL with imprecise parts [52], [54], [55]. SKOSRec builds upon this ideas by adding further expressiveness. The central contributions of the new language are summarized below. • Recommendations for an input profile. Whereas the authors in [54] propose to generate recommendations based on RDF datasets that contain both user and item data, it is likely to argue that an e-commerce retailer will not integrate local customer information with Linked Data. The retailer might avoid an integrated solution because of privacy concerns and additional costs and would rather prefer to obtain

recommendations from external repositories. Therefore the SKOSRec query language can formulate recommendation requests for an input profile (see Q1-Q4). • Graph pattern matching for preference information. Preferences can be expressed for variables that are contained in graph patterns (see Q3). • Subquerying with recommendation results. In some areas it might be helpful to reuse recommendation results as a SPARQL-like subquery. Thus, triple stores can be powerfully navigated (see Q2-Q4). The following examples show the applicability of the above presented ideas. Queries conform to the syntax of the SKOSRec query language. Recommendation results were generated from the DBpedia 3.9 dataset. Q1: Conditional recommendations. PREFIX dct: . PREFIX skos: . PREFIX dbc: .

RECOMMEND ?movie TOP 3 PREF WHERE { ?movie dct:subject ?c . ?concept skos:broader* dbc:Christmas_films .}

Input (, ) The Devil Wears Prada Bridget Jones’s Diary The Terminator Raiders of the Lost Ark

Output Love Actually, The Family Stone Scrooge Ben-Hur Die Hard Trancers

Table 5. Query Results (Q1)

Q2: Roll-up recommendations. PREFIX rdf: . PREFIX yago: . PREFIX dbr: .

SELECT DISTINCT ?place (count(?place) as ?count) WHERE { ?sight ?locatedIn ?place . ?place rdf:type yago:City108524735 .} GROUP BY ?place ORDER BY DESC(COUNT(?place)) LIMIT 5 RECOMMEND ?sight TOP 1000 AGG http://dbpedia.org/resource/Berlin PREF ...

Input (…) Checkpoint Charlie East Side Gallery Berlin Wall DDR Museum Stasi Museum Re:publica Berghain Friedrichshain E-Werk Bauhaus Archive

Output Moskau East Berlin Hamburg Trieste Warschau Vancouver London Amsterdam Paris Montreal

Table 6. Query Results (Q2)

Q3: Drill down recommendations. PREFIX . PREFIX .

SELECT DISTINCT ?director ?movie WHERE { ?movie dbo:director ?director . } LIMIT 4 RECOMMEND ?movie TOP 3 AGG dbr:Quentin_Tarantino PREF ?prefMovie BASED ON { ?prefMovie dbo:director dbr:Quentin_Tarantino .}

?director Robert Rodriguez Frank Miller Robert Rodriguez Tony Scott

?movie From Dusk till Dawn Sin City Sin City True Romance

Table 7. Query Results (Q3)

Q4: Cross-domain recommendations. PREFIX PREFIX PREFIX PREFIX PREFIX PREFIX

dct: . skos: . dbc: . rdf: . dbo: . dbr: .

SELECT ?book (COUNT(?book) as ?count) WHERE { ?book dct:subject ?c. ?c skos:broader{,2} dbc:Novels . { SELECT ?book WHERE {?book ?p1 ?o . ?o ?p2 ?band . ?book rdf:type dbo:Book .}}} GROUP BY ?book ORDER BY DESC(COUNT(?book)) LIMIT 3 RECOMMEND TOP 10 PREF dbr:The_Beatles

?book On the road One Flew Over the Cuckoo’s Nest Sometimes a Great Notion Table 8. Query Results (Q4)

6

Conclusion

This paper described how the LOD cloud can be exploited for customized and scalable on-the-fly recommendations. Former query-based recommender systems disregarded the aspects of scalability and data sparsity, whereas many offline LDRS impose restrictions on flexible customizations and frequent updates. Thus, consumers and e-commerce vendors miss out on the opportunity for advanced recommendations. System requirements were derived to meet Linked Data specificities with the purpose to address problems of existing systems and to offer novel recommendation strategies. The SKOSRec prototype implements these requirements by offering a powerful combination of similar resource retrieval and graph pattern matching. Thus, individual and/or business preferences can be flexibly integrated. With the SKOSRec query language at hand, e-commerce retailers could define various query templates that can be adapted to specific usage contexts. For instance, the marketing department could use campaign templates and end users could enter their preferences through a user interface. First evaluation results showed that the presented approach scales up to a large number of triples and achieves meaningful recommendation results (section 5). In the future, it is intended to evaluate the system’s performance more comprehensively on historic datasets and through a user study.

References [1]

[2]

[3]

[4]

[5] [6]

[7]

[8]

[9]

[10]

[11] [12]

[13]

[14]

G. Adomavicius and A. Tuzhilin, “Toward the next generation of recommender systems: A survey of the state-of-the-art and possible extensions,” IEEE Transactions on Knowledge and Data Engineering, vol. 17, no. 6, pp. 734–749, 2005. J. Konstan, B. Miller, D. Maltz, J. Herlocker, L. Gordon, and J. Riedl, “GroupLens: applying collaborative filtering to Usenet news,” Communications of the ACM, vol. 40, no. 3, pp. 77–87, 1997. L. Terveen, W. Hill, B. Amento, D. McDonald, and J. Creter, “PHOAKS: A system for sharing recommendations,” Communications of the ACM, vol. 40, no. 3, pp. 59–62, 1997. U. Shardanand and P. Maes, “Social information filtering: algorithms for automating ‘word of mouth,’” Proceedings of the SIGCHI conference on Human factors in computing systems, pp. 210–217, 1995. M. Balabanović and Y. Shoham, “Fab: content-based, collaborative recommendation,” Communications of the ACM, vol. 40, no. 3, pp. 66–72, 1997. M. Schmachterberg, C. Bizer, and H. Paulheim, “State of the LOD Cloud 2014,” 2014. [Online]. Available: http://linkeddatacatalog.dws.informatik.uni-mannheim.de/state/. [Accessed: 10-Oct-2015] T. Di Noia, R. Mirizzi, V. Ostuni, and D. Romito, “Exploiting the Web of Data in Model-based Recommender Systems,” Proceedings of the 6th ACM Conference on Recommender Systems, pp. 253–256, 2012. T. Di Noia, R. Mirizzi, V. Ostuni, D. Romito, and M. Zanker, “Linked Open Data to Support Content-based Recommender Systems,” Proceedings of the 8th International Conference on Semantic Systems, pp. 1–8, 2012. L. Peska and P. Vojtas, “Enhancing Recommender System with Linked Open Data,” Flexible Query Answering Systems: 10th International Conference, FQAS, pp. 483–494, 2013. S. Harispe, S. Ranwez, S. Janaqi, and J. Montmain, “Semantic Measures Based on RDF Projections: Application to Content-Based Recommendation Systems,” On the Move to Meaningful Internet Systems: OTM 2013 Conferences, pp. 606– 615, 2013. G. Adomavicius and A. Tuzhilin, “Multidimensional recommender systems: a data warehousing approach,” Electronic commerce, pp. 180–192, 2001. G. Koutrika, B. Bercovitz, and H. Garcia-Molina, “FlexRecs: expressing and combining flexible recommendations,” Proceedings of the 2009 ACM SIGMOD International Conference on Management of data, pp. 745–758, 2009. G. Adomavicius, A. Tuzhilin, and R. Zheng, “REQUEST: A query language for customizing recommendations,” Information Systems Research, vol. 22, no. 1, pp. 99–117, 2011. J. Schafer, J. Konstan, and J. Riedl, “E-commerce recommendation applications,” Applications of Data Mining to Electronic Commerce, pp. 115–153, 2001.

[15] A. Felfernig, S. Gordea, D. Jannach, E. Teppan, and M. Zanker, “A short survey of recommendation technologies in travel and tourism,” OEGAI Journal, vol. 25, no. 7, pp. 17–22, 2007. [16] C. Unger, L. Bühmann, J. Lehmann, A.-C. Ngomo, D. Gerber, and P. Cimiano, “Template-based question answering over RDF data,” Proceedings of the 21st international conference on World Wide Web, pp. 639–648, 2012. [17] C.-N. Ziegler, “Semantic web recommender systems,” Current Trends in Database Technology-EDBT 2004 Workshops, pp. 78–89, 2005. [18] M. Morsey, J. Lehmann, S. Auer, C. Stadler, and S. Hellmann, “DBpedia and the live extraction of structured data from wikipedia,” Program: Electronic Library and Information Systems, vol. 46, no. 2, pp. 157–181, 2012. [19] World Wide Web Consortium, “RDF 1.1 Concepts and Abstract Syntax,” 2014. [Online]. Available: http://www.w3.org/TR/rdf11-concepts/. [Accessed: 11-Oct2015]. [20] N. Marie, F. Gandon, M. Ribière, and F. Rodio, “Discovery Hub: On-the-fly Linked Data Exploratory Search,” Proceedings of the 9th International Conference on Semantic Systems, pp. 17–24, 2013. [21] S. Harris, A. Seaborne, “SPARQL 1.1 query language,” W3C Recommendation, 2013. [Online]. Available: http://www.w3.org/TR/sparql11-query/. [Accessed: 10-Oct-2015] [22] J. Pérez, M. Arenas, and C. Gutierrez, “Semantics and complexity of SPARQL,” ACM Transactions on Database Systems (TODS), vol. 34, no. 3, p. 16, 2009. [23] Ò. Celma and X. Serra, “FOAFing the music: Bridging the semantic gap in music recommendation,” Web Semantics, vol. 6, no. 4, pp. 250–256, 2008. [24] B. Mobasher, X. Jin, and Y. Zhou, “Semantically enhanced collaborative filtering on the web,” Web Mining: From Web to Semantic Web, pp. 57–76, 2004. [25] I. Fernández-Tobías, I. Cantador, M. Kaminskas, and F. Ricci, “A generic semantic-based framework for cross-domain recommendation,” Proceedings of the 2nd International Workshop on Information Heterogeneity and Fusion in Recommender Systems, pp. 25–32, 2011. [26] O. Ozdikis, F. Orhan, and F. Danismaz, “Ontology-based Recommendation for Points of Interest Retrieved from Multiple Data Sources,” Proceedings of the International Workshop on Semantic Web Information Management, pp. 1–6, 2011. [27] M. Stankovic, J. Jovanovic, and P. Laublet, “Linked Data Metrics for Flexible Expert Search on the Open Web,” Proceedings of the 8th Extended Semantic Web Conference (ESWC), pp. 108–123, 2011. [28] V. Ostuni, T. Di Noia, E. Di Sciascio, and R. Mirizzi, “Top-n recommendations from implicit feedback leveraging linked open data,” Proceedings of the 7th ACM conference on Recommender systems, pp. 85–92, 2013. [29] J. Waitelonis and H. Sack, “Towards exploratory video search using linked data,” Multimedia Tools and Applications., vol. 59, no. 2, pp. 645–672, 2011. [30] A. Calì, S. Capuzzi, M. Dimartino, and R. Frosini, “Recommendation of Text Tags in Social Applications Using Linked Data,” Current Trends in Web Engineering, pp. 187–191, 2013.

[31] Y. Hu, Z. Wang, W. Wu, J. Guo, and M. Zhang, “Recommendation for Movies and Stars Using YAGO and IMDB,” Proceedings of the 12th International AsiaPacific Web Conference (APWEB), pp. 123–129, 2010. [32] T. Bannwart, A. Bouza, G. Reif, and A. Bernstein, “Private Cross-page Movie Recommendations with the Firefox add-on OMORE,” Proceedings of the 8th International Semantic Web Conference (ISWC), 2009. [33] T. Ruotsalo, K. Haav, A. Stoyanov, S. Roche, E. Fani, R. Deliai, E. Mäkelä, T. Kauppinen, and E. Hyvönen, “SMARTMUSEUM: A mobile recommender system for the Web of Data,” Web Semantics, vol. 20, pp. 50–67, 2013. [34] S. Gordea, A. Lindley, and R. Graf, "Computing Recommendations for Long Term Data Accessibility basing on Open Knowledge and Linked Data," Joint Proceedings of the RecSys 2011 Workshops Decisions@RecSys’11 and UCERSTI 2, 2011. [35] E. Mannens, S. Coppens, T. Pessemier, H. Dacquin, D. Deursen, R. Sutter, and R. de Walle, “Automatic news recommendations via aggregated profiling,” Multimedia Tools and Applications, vol. 63, no. 2, pp. 407–425, 2011. [36] I. Celino and D. Cerizza, “Anatomy of a Semantic Web-enabled Knowledgebased Recommender System,” Proceedings of the 4th International Workshop Semantic Matchmaking and Resource Retrieval in the Semantic Web, at the 9th International Semantic Web Conference (ISWC), 2010. [37] H. Khrouf and R. Troncy, “Hybrid Event Recommendation Using Linked Data and User Diversity,” Proceedings of the 7th ACM Conference on Recommender Systems, pp. 185–192, 2013. [38] M. Stankovic, W. Breitfuss, and P. Laublet, “Discovering Relevant Topics Using DBPedia: Providing Non-obvious Recommendations,” Proceedings of the 2011 IEEE/WIC/ACM International Conferences on Web Intelligence and Intelligent Agent Technology, pp. 219–222, 2011. [39] J. Golbeck and J. Hendler, “Filmtrust: Movie recommendations using trust in web-based social networks,” Proceedings of the IEEE Consumer Communications and Networking Conference, pp. 282–286, 2006. [40] B. Varga and A. Groza, “Integrating DBpedia and SentiWordNet for a tourism recommender system,” Proceedings of the 2011 IEEE International Conference on Intelligent Computer Communication and Processing (ICCP), pp. 133–136, 2011. [41] A. Passant, “dbrec — Music Recommendations Using DBpedia,” Proceedings of the 9th International Semantic Web Conference (ISWC). pp. 209–224, 2010. [42] M. Fossati, C. Giuliano, and G. Tummarello, “Semantic Network-driven News Recommender Systems: a Celebrity Gossip Use Case.,” International Workshop on Semantic Technologies meet Recommender Systems & Big Data, at the 11th International Semantic Web Conference (ISWC), pp. 25–36, 2012. [43] J. Ahn and X. Amatriain, “Towards Fully Distributed and Privacy-Preserving Recommendations via Expert Collaborative Filtering and RESTful Linked Data,” Proceedings of the 2010 IEEE/WIC/ACM International Conference on Web Intelligence and Intelligent Agent Technology (WI-IAT), pp. 66–73, 2010.

[44] A. Wardhana and H. Nugroho, “Combining FOAF and music ontology for music concerts recommendation on facebook application,” Proceedings of the Conference on New Media Studies (CoNMedia), pp. 1–5, 2010. [45] A. Passant and Y. Raimond, “Combining Social Music and Semantic Web for music-related recommender systems,” Proceedings of the 1st Workshop on Social Data on the Web (SDoW2008), 2008. [46] M. Kaminskas, I. Fernández-Tobías, F. Ricci, and I. Cantador, “Knowledgebased Music Retrieval for Places of Interest,” Proceedings of the 2nd International ACM Workshop on Music Information Retrieval with User-centered and Multimodal Strategies, pp. 19–24, 2012. [47] F. Zarrinkalam and M. Kahani, “A multi-criteria hybrid citation recommendation system based on linked data,” 2nd International eConference on Computer and Knowledge Engineering (ICCKE), pp. 283–288, 2012. [48] R. Meymandpour and J. Davis, “Recommendations Using Linked Data,” Proceedings of the 5th Ph.D. Workshop on Information and Knowledge, pp. 75–82, 2012. [49] Y. Kabutoya, R. Sumi, T. Iwata, T. Uchiyama, and T. Uchiyama, “A Topic Model for Recommending Movies via Linked Open Data,” 2012 IEEE/WIC/ACM International Conferences on Web Intelligence and Intelligent Agent Technology (WI-IAT), pp. 625–630, 2012. [50] V. Groues, Y. Naudet, and O. Kao, “Adaptation and Evaluation of a Semantic Similarity Measure for DBpedia: A First Experiment,” 7th International Workshop on Semantic and Social Media Adaptation and Personalization (SMAP), pp. 87–91, 2012. [51] B. Heitmann and C. Hayes, “Using Linked Data to Build Open, Collaborative Recommender Systems.,” AAAI Spring Symposium: Linked Data meets Artificial Intelligence, pp. 76–81, 2010. [52] S. Policarpio, S. Brunk, and G. Tummarello, “Implementation of a SPARQL Integrated Recommendation Engine for Linked Data with Hybrid Capabilities,” Proceedings of the Artificial Intelligence meets the Web of Data (AImWD) Workshop, at the European Conference on Artificial Intelligence, 2012. [53] A. Musetti, A. Nuzzolese, F. Draicchio, V. Presutti, E. Blomqvist, A. Gangemi, and P. Ciancarini, “Aemoo: Exploratory search based on knowledge patterns over the semantic web,” Semantic Web Challenge, 2012. [54] V. Ayala, M. Przyjaciel-Zablocki, T. Hornung, A. Schätzle, and G. Lausen, “Extending SPARQL for Recommendations,” Proceedings of the 6th International Workshop on Semantic Web Information Management, 2014. [55] C. Kiefer, A. Bernstein, and M. Stocker, “The fundamentals of isparql: A virtual triple approach for similarity-based semantic web tasks. Proceedings of the 6th International Semantic Web Conference (ISWC), 2007.

[56] A. Hajra, A. Latif, and K. Tochtermann, “Retrieving and ranking scientific publications from linked open data repositories,” Proceedings of the 14th International Conference on Knowledge Technologies and Data-driven Business, 2014.

[57] A. Isaac and E. Summers (eds.), “SKOS - Simple Knowledge Organization System Primer,” 2008. [Online]. Available: http://www.w3.org/TR/skos-primer/. [Accessed: 11-Oct-2015]. [58] S. Bechhofer and A. Miles (eds.), “SKOS - Simple Knowledge Organization System Reference,” W3C Recommendation, 2009. [Online]. Available: http://www.w3.org/TR/skos-reference/. [Accessed: 11-Oct-2015]. [59] “SKOS Datasets,” 2015. [Online]. Available: http://www.w3.org/2001/sw/wiki/SKOS/Datasets. [Accessed: 11-Oct-2015].