SemRank: Ranking Complex Relationship Search Results on the Semantic Web Kemafor Anyanwu, Angela Maduko, Amit Sheth LSDIS lab, University of Georgia Paper presentation at WWW2005, Chiba Japan Kemafor Anyanwu, Angela Maduko, and Amit Sheth. SemRank: Ranking Complex Relationship Search Results on the Semantic Web, Proceedings of the 14th International World Wide Web Conference (WWW2005), Chiba, Japan, May 10-14, 2005, pp. 117-127. This work is funded by NSF-ITR-IDM Award#0325464 titled ‘SemDIS: Discovering Complex Relationships in the Semantic Web’ and NSF-ITR-IDM Award#0219649 titled ‘Semantic Association Identification and Knowledge Discovery for National Security Applications.’ Outline • The Problem • The SemRank relevance model • SemRank computational issues in the SSARK system • Evaluating SemRank: strategy and issues • Related Work • Conclusion and Future work The Problem • [Anyanwu et al WWW2003] proposed a query operator for finding complex relationships between entities • [Angles et al ESWC05] a survey of graphbased query operations that should be enabled on the Semantic Web • Question: How can results of relationship query operations be ranked? The Relationship Ranking Problem 8 query q = (1, 3) (a pair of nodes) h g a 1 b d 4 g 2 1. 1 b f d 4 g e f c 6 5 7 1. Find the subgraph that covers q 2. List the results in order of relevance 2. 1 could be done with step 1 or 3. as a separate step . . . 2n. Things to think about • Relevance as best match vs. ???? • Homogenous (hyperlinks) vs. heterogeneous relationships • Should relevance be fixed for all situations? • Size of result set potentially large The SemRank Model SemRank’s Design Philosophy • Tenet 1: Thou shall support variable rankings • Tenet 2: Thou must not burden the user with complex query specification • Tenet 3: Thou shall support main stream search paradigms SemRank’s Key Concepts • Modulative Ranking • Relevance: Search Mode + Predictability • Refraction Count – How varied is the result from what is expected from schema? • Information Gain – How much information does a user gain by being informed about a result? • S-Match – Best semantic match with user need (if provided) Low Information Gain Low Refraction Count High S-Match High Information Gain High Refraction Count High S-Match adjustable search mode Modulative Rank Function • Typical preference or rank function – Ranki = wij * attrij • What we want is, given – µ - weight function parameter – and attributes attr1, attr2 … attrk e.g. length – for each attribute, select appropriate weight functions from g1, g2, … gm e.g. gi (µ) = µ • each gi is some function of µ • Then – Ranki(µ ) = gj(µ) * (attrik) • where gj is the weight function selected for attrk Refraction as a measure of predictability Refraction Student 1 enrolled_in enrolled_in Spouse Course 2 taught_by taught_by Professor 3 married_to 4 • The path “ enrolled_in taught_by married_to “ doesn’t exist anywhere at schema layer • We say that the path refracts at node 3 • High refraction count in a path low predictability married_to Semantic Summary p Representative Ontology Class 3 p1 p2 C1 p1, p2 C3 p1 p2 p5 C2 p4 p5 p1, p2 C5 p4 C4 p3 C1 C1 C2 p1, p2, C3 C4 p 1 , p2 C2 C3 C4 C5 C1 C3 C2 C4 C5 p3 P5, p4 p 1 , p2 p1, p2 p3 p4, p5 Semantic Summary & Refraction. • A Semantic Summary is a graph of representative ontology classes with appropriate relations as arcs • For a path p = r1, p1, r2, p2, r3, there is a refraction at r2 if • p1 (ROCi, ROCj) and p2 (ROCj, ROCk) (or vice versa) where – ROCi, ROCj, ROCk are representative ontology classes of r1, r2, r3 respectively Information content and Information gain Measuring Information Content of a Property • Content is related to uncertainty removed • Typically measured as some function of its probability – High probability -> low information content • For p P, P = set of property types, its information content ISP can be measured as: – ISP(pk) = log2(1/Prk(p = pk)) = - log2 ( [[ pk]] / [[ P ]] ) • ISP(p) is maximum when – Pri = 1 / [[ P ]] = log [[ P ]] Information Content of a Property Sequence – global perspective • The information content of a sequence of properties p1 p2 p3 pk is – max(ISP(pi)), 1 ≤ i ≤ k weak point p1 p2 Prob = high Prob = low p3 Prob = high Information content is dependent on p2 Information Content – Local Perspective • Global high information content but local low information content • Given (a, p1, b), information content with respect to only the valid possibilities between a and b ? (a, p1, b), and valid(p1) is P = (ROCi, ROCj), a ROCi and b ROCj and superproperties • Recompute probabilities based on P (local) – I =min(NI(pi) + average of other NI Total Information Content Total information content = Information content from global perspective + Information content from local perspective S-Match Relevance Specification as keywords Keywords published_in located_in S-Match • Uses the “best semantic match” paradigm • For a keyword ki and a property pj on a path: – SemMatch(ki, pj) = 0 < (2d)-1 1, where • d is the minimum distance between the properties in a property hierarchy • For a path ps, its S-Match value is: – the sum of the max(SemMatch(ki, pj)) Putting it all together ……. SemRank • For a search mode and a path ps: • Modulated information gain for ps, I(ps) – I(ps) = (1-)(I(ps))-1 + I(ps) • Modulated Refraction Count RC(ps) – RC(ps) = RC(ps) • SEMRANK(ps) = I(ps) (1+RC(ps)) (1+S-Match(ps)) Computing SemRank in SSARK The SSARK system Preprocessing phase Query Processing phase Ranking phase x ?? ?? ?? y RDF Documents Loader Preprocessor Storage Manager LtStore UtStore User SubSystem Query & Result Interface Query Processor LAC Look Ahead Cache RC Result Cache Pipelined top-k results Index Manager FDIX PHIX ROIX Ranking Engine Approach g a 1 b 2 d 4 Query Processor f e g,3 c 5 Ranking engine b, 4 Assigns SemRank* values to leaves of the tree i.e. edges on the path * - without refraction count d,6 b ,4 a ,2 c ,1 e ,2 f, 5 f, 5 The Index Subsystem • FDIX – Frequency Distribution IndeX – Stores the frequency distribution of properties • ROIX – Representative Ontology IndeX – Maps classes to Representative Ontology Classes – Stores the semantic summary graph • PHIX – Property Hierarchy IndeX – Uses the Dewey Decimal labeling scheme to encode the hierarchical relationships in a property hierarchy – Used for computing S-Match (match between keywords and properties in a path) ∙ ∙ a, 3 ∙ d, 1 , c.f, 9 g.i, 9 cf, 9 ce, 6 , h, 1 i, 6 a, 3 ∙ , cf, 9 ce , 6 c 4 d, 1 , gi, 9 gh, 4 g, 3 , i, 6 h, 1 ∙ g, 3 , c, 4 f, 5 h, 1 b, 2 c4 d1 , c, 4 d, 1 e, 2 f, 5 e, 2 d, 1 e, 2 f, 5 f, 5 i, 6 , , i, 6 h, 1 , e, 2 ∙ a, 3 ∙ g.i, 9 , i, 6, gh, 4 ∙ h, 1 g, 3 h, 1 i, 6 , cf, 9 , f, 5, ce, 6 e, 2 h, 1 i, 6 ∙ b, 2c, 4 d, 1 e, 2 f, 5 ab, 5a, 3 ∙ c, 4 d, 1 f, 5 ∙ ab, 5 e, 2 b, 2 ∙ ab, 5 g, 3 b, 2 c, 4 Top-K Evaluation Final Top_k: 1. g.i, 18 2. c. f, 9 Evaluation Issues • Data set needs – Entities described with a variety of relationships – Richly connected hierarchies – Realistic frequency distributions • Synthetically generated realistic small data set using human defined rules – e.g. |(p = “audits”)| ≤ 0.1 |(p = “enrolls”)| µ=0 µ=1 Related Work • Semantic searching and ranking of entities on the Semantic Web • Rocha et al WWW2004, Guha et al WWW2003, Stojanovic et al ISWC 2003 , Zhuge et al WWW2003, • Semantic ranking of relationships • Halaschek VLDB demo 2004, Aleman-Meza et al SWDB03 Future Work • Comprehensive evaluation • Including some measures for importance of nodes in the paths • Revise the Modulation function • Optimizing Top-K evaluation – Decreasing height of tree – estimation techniques for a closer approximation to SemRank ordering Data, demos, more publications at SemDis project web site (Google: semdis) Thank You