User Interests Driven Web Personalization based on Multiple Social Networks

Transcription

1 User Interests Driven Web Personalization based on Multiple Social Networks Yi Zeng, Ning Zhong, Xu Ren, Yan Wang International WIC Institute, Beijing University of Technology P.R. China

2 Semantic Data at Web Scale From large scale Web pages to large scale linked open semantic data Number of Web Pages that Google indexes 1998: 270 million 2000: 1 billion 2008: 1 trillion March, 2010: 13 Billion RDF Triples October, 2011: 31.6 Billion RDF Triples June, 2011: 12 Billion RDF Triples from the Web Linking Open Data cloud diagram, by Richard Cyganiak and Anja Jentzsch.

3 The Large Knowledge Collider (LarKC) Project 11 Countries 13 Research Institutions and Universities 3

4 Personalization for Large scale and Web Enabled Semantic Data Processing (cont.) An illustration of the basic idea: Original datasets (Semantic Web Dog Food, Twitter, SwetoDBLP) [s, p, semantic Web mining ] Selected triple set that are related to user interests Interests analysis, evaluation and ranking Frank van Harmelen s Ranked Interests Interests related triples Semantic Spyros Kotoulas RDF Ivan Herman [s, p, RDF triple store ] [s, p, Spyros Kotoulas ] Knowledge DERI For more details: Yi Zeng, Erzhong Zhou, Yan Wang, Xu Ren, Yulin Qin, Zhisheng Huang, Ning Zhong. Research Interests : Their Dynamics, Structures and Applications in Unifying Search and Reasoning. Journal of Intelligent Information Systems, Volume 37, Number 1, 65-88, Springer, Yi Zeng, Ning Zhong, Yan Wang, Yulin Qin, Zhisheng Huang, Haiyan Zhou, Yiyu Yao, and Frank van Harmelen. Usercentric Query Refinement and Processing Using Granularity Based Strategies. Knowledge and Information Systems, Volume 27, Number 3, , Springer, Yi Zeng, Zhisheng Huang, Fenrong Liu, Xu Ren, Ning Zhong. Interest Logic and Its Application on the Web. Proceedings of the 5th International Conference on Knowledge Science, Engineering, and Management (KSEM 2011). Lecture Notes in Artificial Intelligence, Springer, Irvine, California, USA, 2011.

5 Personalization for Large scale and Web Enabled Semantic Data Processing (cont.) A Comparative Study of Query Time and Efficiency for Different Strategies SwetoDBLP dataset: 1.49x10 7 RDF Triples Participants 7 DBLP authors: Preference order 100% : List 2, List 3 List 1 Preference order 100% : List 2 List 3 Preference order 83.3% : List 2 > List 3 List 1 Preference order 16.7% : List 3 > List 2 List 1 See references in the previous page

6 Massive Semantic Data from the Social Web The social Web platforms and the microblog platforms adopt and benefit from semantic techniques The semantic Web gets huge data from these Social Web platforms. 845 million active users Friends Personal Notes Likes Cyber-Social Sensors 150 million users Friends Professional Interests Education Information Work Experiences 350 million users 300 million tweets per day 1.6 billion queries per date Interesting Places Interesting Events 60 million users Following, Followers Real time personal information interesting news From Web of Contents to Web of People Users play more and more important roles

7 Personal Interests Data Fusion Strategies Weighted Fusion Strategy: I( i) = wn I( i) m n= 1 n Average fusion strategy w w n = 1/ n wn = 1 1 w2. Time-sensitive fusion strategy w1 : w2 :...: wn = f 1 : f 2 :...: f n w1 + w wn = 1 Slides 7-10 are from our following paper: Yunfei Ma, Yi Zeng, Xu Ren, and Ning Zhong. User Interest Modeling Based on Multi-source Personal Information Fusion and Semantic Reasoning. Proceedings of the 2011 International Conference on Active Media Technology, Lecture Notes in Computer Science 6890, , Springer, Lanzhou, China, September 7-9, 2011.

8 An Illustration of Multi-source Personal Interests Fusion Evolution of Scientific Information Sharing Open Science Challenges Journal Tradition with Web Collaboration User: Frank van Harmelen Data Source: Interest Value Linked data Open data Web RDF Semantic Web LarKC SPARQL RDFa Science Project Search Engine Symposium A comparative study of interests from three single sources PhD Drupal Information Computer Industry Twitter Facebook LinkedIn Top-K interests from different sources Some of the interests have overlaps among each other. Interest Terms Diversities among these Top-K interests are even more obvious. Research Amsterdam University Educational Institute Knowledge Representation Professor Scientific Director

9 An Illustration of Multi-source Personal Interests Fusion Update frequency: Twitter: f 1 =2.5, Facebook: f 2 =0.2, LinkedIn: f 3 = (per day) Weighted Interests Fusion Function: I ( i) = I ( i) I ( i) I ( i) Interest Values Twitter Average Fusion Time-sensitive Fusion Linked data Open data Web RDF Semantic Web LarKC SPARQL RDFa Science Project Search Engine Interest Terms Symposium PhD A comparative study of interests from a single source and multiple interests sources Average Fusion : Twitter(7) Facebook(7),LinkedIn(2) Time Sensitive Fusion: (1) Top-10 overlaps with Twitter; (2) Values are very close to the ones from Twitter, but entirely different; (3) No interests from Facebook and LinkedIn.

10 Interests Representation and Reasoning about Interests Interests Representation using e-foaf:interest <foaf:person rdf:about=" <foaf:name>frank van Harmelen</foaf:name> <e-foaf:interest> <rdf:description rdf:about=" <dc:title>rdf</dc:title> <e-foaf:cumulative_interest_value rdf:parsetype="resource"> <rdf:value rdf:datatype="&xsd;number"> </rdf:value> </e-foaf:cumulative_interest_value> </rdf:description> </e-foaf:interest>... </foaf:person> A Fragment of AI Ontology ( Frank van Harmelen is interested in RDF in a certain degree RDF representation of AI Ontology <rdfs:class rdf: ID="Graph-based Representation"> <rdfs:subclassof rdf: resource="knowledge Representation"/> </rdfs:class> <rdfs:class rdf: ID="RDF"> <rdfs:subclassof rdf: resource="graph-based Representation"/> </rdfs:class> Reasoning about interests from RDF to Knowledge Representation Appeared on Frank van Harmelen s homepage, but not elsewhere.

11 Active Academic Visit Recommendation Application (AAVRA) Collaboration network is already too complex, but Academic collaboration candidates not only appear on publication data, but also on many other social networking environment such as Twitter. A Snapshot from Semantic Web Dog Food Affiliation Map Data Sources: Twitter Data, Semantic Web Dog Food data, Google Maps API

12 AAVRA: Data Acquisition Twitter data acquisition Twitter data acquisition to : Locate the end user; Find agents that the end user follows; User real time interests analysis; Locating followings and their interests

13 AAVRA: Data Acquisition from SWDF Real time acqusition by SPARQL end point SELECT DISTINCT $person $person_name $affiliation $affiliation_name WHERE { $person a foaf:person. $person foaf:name $person_name. $person foaf:made $InProceedings. $InProceedings foaf:maker $person_url. $person_url foaf:name "Frank van Harmelen". $person swrc:affiliation $affiliation. $affiliation foaf:name $affiliation_name }

14 AAVRA: Generating Levels of Recommendation Interpretations on different groups of data from SWDF and Twitter Interest Levels C o a u t h o r Formula (, ) T F i n g( u, p) Result Sets 1 S WD F p u T 1 C o a u t h o r (, ) T F i n g( u, p) T F i n g( u, p) P C o a u t h o rs WDF p u 2 S WDF p u T 2 3 (, ) T 3 4 T F i n g( u, p) S I T( p, u, K) S WDF( p) T 4 5 T F i n g( u, p) S I T ( p, u, K) S W D F( p) T 5

15 AAVRA: Recommendation Results Analysis Interest Level Recommendati on Ratio(%) Paul Groth Results Examples Spyros Kotoulas(3), Jacopo Urbani(3), Eyal Oren(2), Henri Bal(2), Zharko Aleksovski(2), Zhisheng Huang(1), Kalina Bontcheva, Lynda Hardman, Peter Mika, Steffen Staab, Denny Vrandecic, Ivan Herman, Michael Hausenblas, Stefano Bertolo, Dan Brickley, DERI Galway, Web Foundation, Ontotext AD... Recommendation Ratio = Recommended Results / Candidate Space Candidate Space: 7131 persons (SWDF+Twitter) Calculation of SIT(p,u,K), Top-10 interests, K= % candidates are recommended overall.

16 Active Academic Visit Recommendation: A Snapshot The 3 rd level of recommendation: T F i n g( u, p) P C o a u t h o rs WD F p u University of Sheffield (Kalina Bontcheva) University of the West of England (Richard McClatchey) (, )

17 Into the Future A conservative estimate would be that it would take 10,000 triples just to describe each human, which gives us 100 trillion (10 14 ). Pictures from Prof. Ning Zhong s plenary talk at Web Intelligence 2011

18 Thank You!