Health Articles

 

Dojchinovski.mk

Personalised, Serendipitous and Diverse Linked Data Resource Recommendations Milan Dojchinovski and Tomas Vitvar Web Intelligence Research Group Faculty of Information Technology Czech Technical University in Prague Abstract. Due to the huge and diverse amount of information, the ac-tual access to a piece of information in the Linked Open Data (LOD)cloud still demands significant amount of effort. To overcome this prob-lem, number of Linked Data based recommender systems have been de-veloped. However, they have been primarily developed for a particulardomain, they require human intervention in the dataset pre-processingstep, and they can be hardly adopted to new datasets. In this paper, wepresent our method for personalised access to Linked Data, in particularfocusing on its applicability and its salient features.
Keywords: personalisation, recommendation, Linked Data, semanticdistance, similarity metric Due to the huge and diverse amount of information, the actual access to a pieceof information in the Linked Open Data (LOD) cloud still demands significantamount of effort. To overcome this problem, number of Linked Data based rec-ommender systems have been developed. However, they have been primarilydeveloped for a particular domain (e.g., music or videos), they require humanintervention in the dataset pre-processing step, and they can be hardly adoptedto other datasets.
In our related EKAW 2014 research paper we described in detail our method for "personalised access to Linked Data". In this paper, we focus onits usage in different domains and datasets, and its salient features. Section describes the resource recommendation workflow. Section presents a resourcerecommendation case study from the Web services domain. Finally, Section highlights its salient features.
Linked Data Resource Recommendation The prototype of our method implements the workflow depicted in Figure Itstarts with (i) Linked Data datasets and user profiles import, followed by (ii) Milan Dojchinovski and Tomas Vitvar Resource Similarity
Resource Relevance
Personalised Resource
RDF Datasets Import
Linked Open Data
Users' Profiles
Fig. 1. Overview of the resource recommendation workflow graph based data analysis, and ends with (iii) generation of personalised LinkedData resource recommendations.
RDF Datasets Import. First, it is necessary to identify one or more datasetswhich contain the resources of interest. The datasets can be single domaindatasets such as GeoNames, or multi-domain datasets such as DBpedia. It isalso necessary to provide users' profiles information which encodes users' activi-ties and their relations to other people and resources. According to the statisticsas of April 2014 48% of all datasets in the LOD cloud are datasets providingsocial information such as people profiles data and social relationships amongthem. Note that the RDF data is imported and analysed in its original form. Inother words, we do not perform any domain-dependent selection or filtering ofthe information. For comparison, the movie recommender described in usesonly a subset of the information, defined as a set of RDF properties.
Graph Analysis. Our resource recommendation method is based on the col-laborative filtering technique and it recommends resources from users with in-terests in similar resources. To measure the user similarity, we perform graphbased analysis of the users' RDF context graphs. To this end, we introduce anovel domain–independent semantic resource similarity metric, which takes intoaccount the commonalities and informativeness of the shared resources. Ourassumption is that the more information two resources share, the more similarthey are. And second, the more informative resources they have in common, themore similar they are. The resources informativeness is primarily incorporated todifferentiate informative shared resources from non-informative (e.g., owl:Thingor skos:Concept ).
Figure illustrates computation of similarity between two users, Alfredo and mlachwani. The users have 6 resources in common, which convey differentamount of information content. According to our assumption, resources with highgraph degree value convey less information content, than resources with lowergraph degree. Considering the whole Linked Web APIs dataset, the FacebookAPI with node degree of 418, is more informative and carries more similarityinformation than the Twitter API, with node degree of 799. Facebook API ischaracterised with a lower degree due to its lower usage in mashups, leading toa lower number of incident links.
Personalised Resource Recommendations. The similarities between theresources representing users are then used to compute the relevance of eachresource candidate. For a user requester, the method first computes the con- Linked Data Resource Recommendations Fig. 2. Excerpt from the Linked Web APIs dataset with resource context graphs withcontext distance of 3 nectivity between each similar user (i.e., a user similar with the user requester)and each resource candidate. The connectivity is computed as the amount ofgained information between the similar user and the resource candidate. Thefinal relevance score aggregates the user similarity scores and the connectivityscores for the resource candidate. Finally, the method returns the top-n mostrelevant resources for the user requester.
Case Study: Recommendation of Linked Web APIs In this section we show a resource recommendation case study on the Linked WebAPIs dataset which provides RDF representation of the ProgrammableWeb.comservice repository as of April 24th 2014. The datasets provides descriptions for11,339 Web APIs, 7,415 mashups and 5,907 users.
Step 1: Identification of requesters and items of interest. We startwith identification of resources representing requesters and resources represent-ing items of interest. In our case, requesters are users represented with thefoaf:Person class and items of interests are Web APIs represented with thewl:Service class. If we want to develop a music recommendation system forrecommendation of artists and bands, it will be necessary to specify the classdescribing those resources of interest. In DBpedia those are resources of typedbpedia:MusicalArtist and dbpedia:Band.
Step 2: Data analysis. Next, the method computes the similarities betweenresources representing users (i.e., foaf:Person) and computes the relevance ofeach resource candidate for each user. In this step, the resource informationcontent is also computed and considered.
Step 3: Generate resource recommendations. Finally, the method cangenerates a list of top-n most relevant resources for the user requester.
Milan Dojchinovski and Tomas Vitvar Serendipitous and Diverse Recommendations. For any recommender sys-tem it is important to accommodate the difference between the individuals inorder to produce accurate, while at the same time serendipitous and diverse rec-ommendations. The results from our evaluation show that our method satisfythese requirements and also outperforms the traditional personalised collabora-tive filtering methods and non-personalised methods.
Adaptability. One drawback of the existing Linked Data based recommenda-tion systems is their applicability on different datasets due to the metricused for computation of the resource similarity. The metrics used in are notsuitable for computation of similarities between resources in datasets (i.e., theLinked Web APIs dataset), where the graph distance between the resources ismore than two. In comparison, our method can be easily adapted to any datasetby setting the size of the resource context. In datasets, where the users in theRDF graph are close to each other, will require setting lower distance, whilein datasets where the users are far, will require higher distance. Choosing smallcontext distance in datasets where the users are far from each other, can possiblylead to no overlap of the resource context graphs, and no similarity evidenced.
Cross-dataset and domain recommendations. As shown in Section with-out any restrictions our method can consume and benefit from one or moreLinked Data datasets from different domains, and at the same time produce re-source recommendations for these domains and datasets. In contrast, the meth-ods presented in use only subsets of Linked Data datasets, which need tobe defined in advance for the particular domain.
Acknowledgement. This work was supported by the Grant Agency of the Czech Technical University in Prague, grant No. SGS14/104/OHK3/1T/18. Wealso thank to Programmableweb.com for supporting this research.
1. T. Di Noia, et al. Linked open data to support content-based recommender sys- tems. In Proceedings of the 8th International Conference on Semantic Systems,I-SEMANTICS '12, pp. 1–8. ACM, New York, NY, USA, 2012.
2. M. Dojchinovski, J. Kuchar, T. Vitvar, and M. Zaremba. Personalised graph-based selection of web apis. In P. Cudr-Mauroux, et al., (eds.) The Semantic Web ISWC2012, vol. 7649 of Lecture Notes in Computer Science, pp. 34–48. Springer BerlinHeidelberg, 2012.
3. M. Dojchinovski and T. Vitvar. Personalised access to linked data. In K. Janow- icz, S. Schlobach, P. Lambrix, and E. Hyvnen, (eds.) Knowledge Engineering andKnowledge Management, vol. 8876 of Lecture Notes in Computer Science, pp. 121–136. Springer International Publishing, 2014.
4. A. Passant. dbrec music recommendations using dbpedia. In P. Patel-Schneider, et al., (eds.) The Semantic Web ISWC 2010, vol. 6497 of Lecture Notes in ComputerScience, pp. 209–224. Springer Berlin Heidelberg, 2010.
5. M. Schmachtenberg, C. Bizer, and H. Paulheim. Adoption of the linked data best practices in different topical domains. In P. Mika, et al., (eds.) The Semantic WebISWC 2014, vol. 8796 of Lecture Notes in Computer Science, pp. 245–260. SpringerInternational Publishing, 2014.

Source: http://www.dojchinovski.mk/public/data/EKAW2014p-DojchinovskiV.pdf