TermPicker: Enabling the Reuse of Vocabulary Terms by Exploiting Data from the Linked Open Data Cloud
- 3. Term recommenda:ons based on…? 3 § rdfs:domain, rdfs:range, and other informa;on encoded in vocabularies? § Popularity of a vocabulary term? § Classes and proper;es from domain specific vocabularies? § etc. Which vocabulary terms did other data providers on the LOD cloud use in a similar scenario?
- 4. How to capture a Scenario? 4 § A scenario is defined by vocabulary terms used for a part of a model à paKerns on schema level Example: slp = ({swrc:Publication}, {dc:creator}, {foaf:Person}) Resources of type swrc:PublicaFon are connected to resources of type foaf:Person via the property dc:creator slp = (sts, ps, ots) Schema-Level PaIerns (SLPs) A tuple describing the connec;on between two sets of classes via a set of proper;es In General:
- 5. Vocabulary Term Recommenda:ons Based on LOD 5 Recommender of vocabulary terms:{x1, ..., xn} query input I query-SLP: slpq = ({mo:SoloMusicArtist}, ?, ?) Feature Computation {F(slpq, x1), ..., F(slpq, xn)}{F(slpq, x1), ..., F(slpq, xn)} II Ranking Model III %({F(slpq, x1), ..., F(slpq, xn)}) query output IV Classes for subject: properties: Clases for object:<..., mo:Record, mo:MusicGroup,...> <..., mo:MusicArtist, foaf:Person,...> <..., foaf:made,..., mo:member of,...>
- 6. Overview 6 Feature Computation {F(slpq, x1), ..., F(slpq, xn)}{F(slpq, x1), ..., F(slpq, xn)} Recommender of vocabulary terms: Ranking Model query input query output IV I II III Classes for subject: properties: Clases for object: query-SLP: {x1, ..., xn} %({F(slpq, x1), ..., F(slpq, xn)}) <..., mo:Record, mo:MusicGroup,...> slpq = ({mo:SoloMusicArtist}, ?, ?) <..., mo:MusicArtist, foaf:Person,...> <..., foaf:made,..., mo:member of,...>
- 7. Feature Computa:on: The SLP-Feature 7 slpq = ({mo:SoloMusicArtist}, {}, {}) If slpq ✓ slpi (slpi 2 SLPLOD) Then Sets of recommendations: slpi slpq Collabora;ve Filtering Classes for subject: < mo:MusicArtist, dbo:Actor > Properties: < mo:member of, foaf:made, mo:recorded > Classes for object: < mo:MusicBand, mo:Record > SLPLOD = {({mo:SoloMusicArtist, mo:MusicArtist}, {mo:member of}, {mo:MusicBand}) ({mo:SoloMusicArtist, dbo:Actor}, {foaf:made, mo:recorded}, {mo:Record}) ({foaf:Person}, {foaf:knows}, {foaf:Person}) } SLPLOD = SPLs computed from datasets on the LOD cloud
- 8. Feature Computa:on: State of the Art Features1 8 Feature Definition of the Feature f1 Number of datasets on the LOD cloud using the recommendation candidate x f2 Number of datasets on the LOD cloud using the vocabulary Vx of recommendation candidate x f3 Total number of occurrences of recommendation candidate x on the LOD cloud f4 Whether the recommendation candidate x is from a vocabulary that is already used in query-SLP slpq f1 f3: Reusing popular vocabularies/vocabulary terms f4: Reusing vocabulary terms from the same vocabulary 1) Schaible, GoKron, and Scherp: Survey on Common Strategies of Vocabulary Reuse in Linked Open Data Modeling (ESWC 2104)
- 9. Overview 9 Feature Computation {F(slpq, x1), ..., F(slpq, xn)}{F(slpq, x1), ..., F(slpq, xn)} Recommender of vocabulary terms: Ranking Model query input query output IV I II III Classes for subject: properties: Clases for object: query-SLP: {x1, ..., xn} %({F(slpq, x1), ..., F(slpq, xn)}) <..., mo:Record, mo:MusicGroup,...> slpq = ({mo:SoloMusicArtist}, ?, ?) <..., mo:MusicArtist, foaf:Person,...> <..., foaf:made,..., mo:member of,...>
- 10. Calcula:ng a Ranking Model 10 How to weight the feature values? § Learning to Rank (L2R): } Family of supervised machine learning algorithms based on data with relevance annota;ons } state of the art in IR to compute a generalized ranking model over a given set of features } Ranking model is derived by observing correla;ons between feature values and candidate relevance F f1 f2 f3 f4 SLP-feature (slpq, x1) 7 9 20 1 4 (slpq, x2) 3 3 5 0 6 (slpq, x3) 10 20 80 0 2 (slpq, x4) 4 20 29 1 4
- 11. Overview 11 Feature Computation {F(slpq, x1), ..., F(slpq, xn)}{F(slpq, x1), ..., F(slpq, xn)} Recommender of vocabulary terms: Ranking Model query input query output IV I II III Classes for subject: properties: Clases for object: query-SLP: {x1, ..., xn} %({F(slpq, x1), ..., F(slpq, xn)}) <..., mo:Record, mo:MusicGroup,...> slpq = ({mo:SoloMusicArtist}, ?, ?) <..., mo:MusicArtist, foaf:Person,...> <..., foaf:made,..., mo:member of,...>
- 12. 12 Evalua:on § Baseline POP: Reuse popular vocabulary terms } Based on features: § Baseline SAME: Reuse terms from same vocabulary } Based on features: § SLP-feature-based: U;lizing the SLP-feature } Based on features: What is the benefit of the SLP-feature? f1 f3 f1 f4 f1 f4 (+ SLP-feature )
- 13. 13 Evalua:on Procedure § Offline evalua;on with hidden informa;on § Measuring quality of recommenda;ons } Mean Average Precision (MAP ) } Mean Reciprocal Rank at the first 5 posi;on (MRR@5) § Use of the RankLib2 library slpq = ({mo:SoloMusicArtist}, {foaf:made}, {mo:Record}) Example: Randomly hidden term: “foaf:made” < foaf:name, mo:remixed, foaf:made, ... >Result list: 2) hKps://sourceforge.net/p/lemur/wiki/RankLib/
- 14. 14 Evalua:on Data for Recommenda:ons § Two evalua;ons based on BTC 20143 and DyLDO4 BTC 2014 DyLDO # of triples first 34 mio. (reduce overhead) 10.8 mio # of PLDs 3, 500 382 # of distinct terms 5.5 mio. 2.3 mio. # of vocabularies 1, 500 600 # of computed SLPs 227, 000 118, 000 § 10-fold leave-one-out valida;on based on PLDs 3) hKp://km.aid.kit.edu/projects/btc-2014/ 4) hKp://swse.deri.org/dyldo/
- 16. 16 Discussion § Using “from same vocabulary”-feature not significant } à only in few cases terms from same vocabulary are used § Using SLPs significant improvement (ca. 35% in MAP) } à already now: looking at how others model their data § BeKer performance on BTC 2014 } More data in BTC 2014 to train the ranking model } 37% more relevant candidates
- 17. 17 Conclusion § Using SLPs, relevant recommenda;ons are ranked significantly higher in the result list } Can aid the engineer even more in modeling data in a way how other data providers do § Using L2R, the more relevant candidates correlate with a feature, the beKer the results However, Offline evalua;ons do not observe actual user behavior à online evalua;on needed5 5) Performed in Schaible, Szekely, Scherp: “Comparing Vocabulary Term Recommenda;ons using Associa;on Rules and Learning To Rank: A User Study” at ESWC 2016