As part of LOD2 project and OpenData.cz initiative, we are developing an ODCleanStore framework enabling management of Linked Data. In this paper, we focus on the query-time data fusion in ODCleanStore, which provides data consumers with integrated views on Linked Data; the fused data (1) has solved con icts according to the preferred con ict resolution policies and (2) is accompanied with provenance and quality scores, so that the consumers can judge the usefulness and trustworthiness of the data for their task at hand.
? The work presented in this article has been funded in part by EU ICT FP7 under No.257943 (LOD2 project), the Czech Science Foundation (GACR, grant number 201/09/H057), and GAUK 3110. 1 http://richard.cyganiak.de/2007/10/lod/ 2 http://opendata.cz, http://lod2.eu 3 To download the code, please visit http://sourceforge.net/p/odcleanstore 4 RDF triples can be extended to quads (s; p; o; g) where g is the named graph [3] to which the data belongs. When talking about \data in the named graph g", we mean all the quads ( ; ; ; g)). application registered in ODCS, e.g. by various extractors. Based on the identier of the feed, the appropriate transforming pipeline is launched; the pipeline successively executes a de ned (and customizable) set of transformers ensuring that data in the processed feed is cleaned, resources deduplicated and linked to already existing resources in the clean database or in the linked open data cloud, data is enriched with new resources, arbitrarily transformed, and the quality of the feed (graph score) is assessed. When the pipeline nishes, the augmented RDF feed is populated to the clean database together with any auxiliary data and metadata created during the pipeline execution, such as links to other resources or metadata about the feed's graph score.
Data consumers can query (via third-party applications) the clean database to obtain data about the certain resource (e.g. a city, such as the German city \Berlin"). Since the same resource can be described by various sources (feeds), con icts may arise when integrating data about that city. To solve this, ODCS applies in the data fusion algorithm certain con ict resolution policies which resolve data con icts in the resulting RDF data; these policies can be customized by the consumer. Furthermore, the resulting integrated RDF data is supplemented with provenance metadata (data origin) and quality scores of the integrated quads, so that data consumers can judge the usefulness and trustworthiness of the resulting data for their task at hand; the quality score is in uenced by the quality of the feed the triples originate from (graph score) and by the applied con ict resolution policy [4]. The data fusion algorithm runs during query time, because consumers in di erent situations can have di erent requirements on the data.
This paper brie y describes the data fusion algorithm in ODCS in Section 1; the algorithm is fully described in [4]. The practical demonstration5 based on the illustrative examples in Section 1 gives further insight into the work of the data fusion algorithm.
To the best of our knowledge, there is just one another linked data fusion software { Sieve { currently under development [5]. Sieve is part of Linked Data Integration Framework6. Di erently from our approach, Sieve fuses data while 5 http://www.ksi.m .cuni.cz/~knap/iswc12 6 http://www4.wiwiss.fu-berlin.de/bizer/ldif/ being stored to the clean database and not during execution of queries, thus, provides no data fusion customization during data querying. 1
Suppose that the clean database of ODCS contains data about the German city Berlin coming from multiple sources { DBpedia, GeoNames, and Freebase7. Let us assume that Alice, a data consumer, is an investigative journalist who is writing a story about Berlin; thus, she submits the keyword \Berlin" to the query execution component of ODCS and she would like to get all the information the framework knows about Berlin fused from the available sources.
When fusing data, the data fusion algorithm in ODCS has to deal with data con icts, which happen when two quads have inconsistent object values for a certain subject s and predicate p; such quads are called o-con icting quads and the con icting object values of these o-con icting quads are called con icting values. The solution of the con icts is prescribed by the con ict resolution policies, which may be speci ed globally or per predicate. We distinguish two types of con ict resolution policies { deciding and mediating. Deciding policies select one or more values from the con icting values, e.g., an arbitrary value (ANY), maximum value (MAX), the value with the highest quality (BEST), or all conicting values (ALL). Mediating policies compute new value, e.g. an average (AVG) of the con icting values. For example, Alice may specify she would like to receive in the response all the distinct values for the subject representing Berlin and predicate rdf:type (deciding con ict resolution policy ALL). On the other hand, she may want to compute for the same subject average value (AVG) for the values of the predicate geo:lat, select the best value with the highest quality (BEST) for rdfs:label of Berlin, and select maximum value (MAX) from the values of the predicate dbprop:populationTotal of Berlin.
When describing the data fusion algorithm within execution of consumer's
queries in ODCS, we suppose that the typical pre-fusing processes [2] { schema
mapping (the detection of equivalent schema elements in di erent sources) and
duplicate detection (detection of equivalent resources) has already been done.
Therefore, we suppose that (
The input to the data fusion algorithm is (
The fusion algorithm rstly replaces URIs of resources representing the same
concept (i.e. connected by an owl:sameAs links) with a single URI and removes
duplicate quads8. Consequently, quads are grouped to the sets of comparable
quads { i.e. quads having the same subject and predicate; o-con icting quads
form subset of the corresponding comparable quads. For each set of comparable
quads, two steps (Step S1 and S2) are executed: Step S1 chooses and applies a
con ict resolution policy determined by the predicate of the comparable quads
and Step S2 computes quality of the quads resulting from Step S1. Multiple
realworld cases lead us to three factors in uencing the computation of the quality
of the resulting fused quads (in Step S2): (
This paper introduces query-time data fusion algorithm in ODCleanStore { the framework for managing Linked Data. The practical demonstration9 shows the maturity of the algorithm and demonstrates its features { application of con ict resolution policies and computation of the quality of the fused quads. Full theoretical background behind the data fusion algorithm is in [4]. 8 Quads having the same subject, predicate, object, and the named graph. 9 http://www.ksi.m .cuni.cz/~knap/iswc12