To allow e ective querying on the Web of data, systems frequently rely on data from multiple sources for answering queries. For instance, a user may wish to combine data from sources comprised in di erent statistical catalogs. Given such federated queries, in order to enable an interactive exploration of results, systems must allow user involvement during data source selection. That is, a user should be able to choose data sources contributing to query results, thereby allowing to re ne/expand current ndings. For this, one needs e ective recommendations for data sources to be picked: data source contextualization. Recent work, however, solely aims at source contextualization for \Web tables", while heavily relying on schema information and simple table structures. Addressing these shortcomings, we exploit work from the eld of data mining and show how to enable e ective Web data source contextualization. Based on a real-world nance use-case, we built a contextualization engine, which we integrated into a Web search system, our data portal, for accessing statistics data sets.
The amount of RDF data available on the Web today, such as Linked Data1, RDFa and Microformats, is large and rapidly increasing.2 RDF data contains descriptions of entities, with each description being a set of triples. A triple associates an entity identi er (subject) with an object via a predicate. A set of triples forms a data graph.
RDF data is oftentimes highly distributed, with each data source comprising one or more RDF graphs (Fig. 1). Most notably, Linked Data as well as Webaccessible SPARQL endpoints have contributed to this development.
Integrated Querying of Multiple Data Sources. In order to ful ll information needs over multiple, distributed data sources, a number of issues need to addressed. These range from the ability to discover and identify relevant data sources, to the ability to integrate them and nally to support querying them in a transparent manner.
Src. 2: gov q ggdebt (Eurostat). es : data / tec0001 r d f : type qb : Observation ; es prop : geo es d i c :DE; es prop : u n i t es d i c :MIO EUR; es prop : i n d i c n a es d i c : B11 ; sd : time "2010 01 01"^^ xs : date ; sm : obsValue " 2496200.0 "^^ xs : double .
es : data / gov q ggdebt r d f : type qb : Observation ; es prop : geo es d i c :DE; es prop : u n i t es d i c :MIO EUR; es prop : i n d i c n a es d i c : F2 ; sd : time "2010 01 01"^^ xs : date ; sm : obsValue " 1786882.0 "^^ xs :
decimal .
Src. 3: NY.GDP.MKTP.CN (Worldbank). wbi :NY.GDP.MKTP.CN r d f : type qb : Observation ; sd : r e f A r e a wbi : c l a s s i f i c a t i o n / country /DE; sd : r e f P e r i o d "2010 01 01"^^ xs : date ; sm : obsValue " 2500090.5 "^^ xs : double ; wbprop : i n d i c a t o r wbi : c l a s s i f i c a t i o n / i n d i c a t o r /NY.GDP.MKTP.CN .
Say a user is interested in economic data. Here, catalogs like Eurostat3 or
Worldbank4, o er rich statistical information about, e.g., GDP, spread across
many sources. However, these data sources are very speci c, and in order to
provide the user with her desired information, a system has to combine data
from multiple sources. Processing queries in such a manner requires knowledge
about what source features which information. This problem is commonly known
as source selection: a system chooses data sources relevant for a given query
and query fragment, respectively. Previous works selected sources by means of
indexes, e.g., [
Data Source Contextualization. Existing approaches for source selection
aim solely at a mapping of queries/query fragments to sources featuring exactly
matching data [
but relevant to her information need. Integration of these additional sources for contextualization of known, relevant sources, provides a user with broader results in terms of result dimensions (schema complement) respectively result entities (entity complement). See also an example in Fig. 1.
For enabling systems to identify and integrate sources for contextualization, we argue that user involvement during source selection is a key factor. That is, starting with an initial search result obtained via, e.g., a SPARQL or keyword query, a user should be able to choose and change sources used for result computation. In particular, users should be recommended contextual sources at each step of the search process. After modifying the selected sources, results may be reevaluated and/or queries expanded.
Recent work on data source contextualization focuses on Web tables [
Contributions. In this work, we provide the following contributions: (1) We present an entity-based solution for data source contextualization in the Web of data. This engine is based on well-known data mining strategies, and does not require schema information or data adhering to a particular form. (2) We implemented our system, the data-portal, based on a real-world use case, thereby showing its practical relevance and feasibility. A prototype version of this portal is freely available and currently tested by a pilot customer.6
Outline. In Sect. 2, we present a real-world use case. In Sect. 3, we outline our contextualization engine, before we discuss the data portal system in Sect. 4. We present related work in Sect. 5. We conclude with Sect. 6. 2
In this section, we introduce a real-world use case to illustrate challenges and opportunities in contextualizing data sources. The scenario is situated in nancial research, provided by a pilot user in a private bank.
In their daily work, nancial researchers heavily rely on a variety of open and closed Web data sources in order to provide prognoses of future trends. A typical example is the analysis of government debt. During the nancial crisis in 20082009, most European countries made high debts. To lower doubts about repaying
these debts, most countries set up a plan to reduce their public budget de cits. The ful llment of these plans is essential for the Euro zone's development.
To analyze such plans, a nancial researcher requires an overview of public revenue and expenditure in relation to the gross domestic product (GDP). To measure this, she needs information about the de cit target, the revenue/expenditure/de cit and GDP estimates. This information is publicly available, provided by catalogs like Eurostat and Worldbank. However, it is spread across a huge space of sources. That is, there is no single source satisfying her information needs, instead data from multiple sources have to be identi ed and combined.
To start her search process, a researcher may give \gross domestic product" as keyword query. The result is GDP data from a large number of sources. At this point, data source selection \hidden" from the researcher, and sources are solely ranked via number and quality of keyword hits. However, knowing where her information comes from is critical. In particular, she may want to restrict and/or know the following meta-data: { General information about the data source, e.g., the name of the author and a short description of the data source contents. { Information about entities contained in the data source, e.g., the single countries of the European Union. { Description about the dimensions of the observations, e.g., the covered time range or the data unit of the observations.
By means of faceted search, the researcher nally restricts her data source to tec00001 (Eurostat, Fig. 1) featuring \Gross domestic product at market prices". However, searching the data source space in such a manner requires extensive knowledge. Further, the researcher was not only interested in plain GDP data { she was also looking for additional information.
For this, a system should suggest data sources that might be of interest, based on sources known to be relevant. These contextual sources may feature related, additional information w.r.t. current search results/sources. For instance, data sources containing information about the GDP of further countries or with a different temporal range. By such means, the researcher may discover new sources more easily, as one source of interest links to another { allowing her to explore the space of sources. 3
In this section, we outline an approach for Web data source contextualization.
For this, we conceive a data source D 2 D as set of multiple RDF graphs,
with D as set comprising all sources in the data space. Further, an entity e is
given by an RDF instance contained in source D, and described by a subgraph
Ge in D [
Related Entities. The intuition behind our approach is simple: if data
sources contain similar entities, they are somehow related. In other words, we
rely on entities to capture the \latent" semantics of data sources. That is, we
employ o ine procedures as follows: we (1) extract entities, (2) measure similarity
between them, and (3) cluster them.
(1) Entity Extraction. We start by extracting entities from each source D. First,
for scalability reasons, we go over all entities in data graphs in D and collect
a sample, with every entity having the same probability of being selected.
For each selected entity e, we crawl its surrounding subgraph { resulting in
a graph Ge that describes e [
Contextualisation Score. Similar to [
The former is an indicator for the entity complement of D00 w.r.t. D0. That
is, ec asks: how many new, similar entities does D00 contribute to given entities
in D0? The latter score, sc, measures how many new \dimensions" are added by
D00, compared to those already present in D0 (schema complement). In contrast
to [
FluidOps Data Portal
Data Source Exploration Data Source
Search
Data Source Visualization
Select/Remove Source for Query Federation
Data Source Contextualization Engine
Inspect Source Contributing to
Current Result Data Access
Data Source
Meta-Data Meta-Data Updates
by Providers Eurostat Provider
Worldbank Provider
Eurostat gov_q_ggdebt tec00001
Entity Clusters
Offline Entity Extraction and
Clustering
Query Processing Sparql Query
Result
Visualization Federation Layer
Query & Visualize Results
Processing SPARQL Queries against the
Federation Loading and Populating of
SPARQL
Endpoints Data Source gov_q_ggdebt
Data Source tec00001
Data Source
NY.GDP.MKTP.CN Data Space
Worldbank
Data Loader NY.GDP.MKTP.CN Fig. 2: The data portal system features two kinds of services: source space exploration, and query processing. For the former, our source contextualization engine is integrated as a key component. Overall, source space exploration requires source meta-data as well as entity clusters to be available. Entity clusters are computed as an o ine process, while meta-data may be updated frequently during runtime. On the other hand, query processing distributes query fragments via a federation layer. Each fragment is evaluated over one or more sources. For this, each data source is mapped to a SPARQL endpoint, for which data is accessed via a data-loader. For our running example, the necessary sources are loaded via three endpoints: gov q ggdebt, tec00001, and NY.GDP.MKTP.CN.
Let us rst de ne an entity complement score ec : D D 7! [0; 1]. In the most simplistic manner, we may measure ec by counting the overlapping clusters between both sources: ec(D00 j D0) :=
X Cj2 cluster(D0) 1(Cj ; D00)jCj j jCj j with cluster as function mapping data sources to clusters their entities are assigned to. Further, let 1(C; D) by an indicator function, returning 1 if cluster C is associated with data source D via one or more entities in D.
Considering the schema complement score, sc : D D 7! [0; 1], we aim to count new dimensions (properties) that are introduced by D00. Thus, a simple realization of sc may be given by: jprops(Cj ) n SCi2 cluster(D0) props(Ci)j jprops(Cj )j with props as function projecting a cluster C to a set of properties, where each property is contained in a description of an entity in C.
Finally, a contextualization score cs is obtained by a monotonic aggregation of ec and sc. In our case, we apply a weighted summation:
cs(D00 j D0) := 1=2 ec(D00 j D0) + 1=2 sc(D00 j D0)
Runtime Behavior and Scalability. Regarding online performance, i.e., computation of contextualization score cs, given the o ine learned clusters, we aimed at simple and lightweight heuristics. For ec only an assignment of data sources to clusters (function cluster(D)), and cluster size jCj is needed. Further, measure sc only requires an additional mapping of clusters to \contained" properties (function props(C)). All necessary statistics are easily kept in memory.
With regard to o ine clustering behavior, we expect our approach to perform
well, as existing work on kernel k-means clustering showed such approaches to
scale to large data sets, e.g., [
We have implemented the presented algorithms (Sect. 3) for data source contextualization in a data-portal, enabling on demand access to data sources from a number of open statistic data catalogs. Based on our real-world use case, we show how the source contextualization is used within this portal.
Overview. Towards an active involvement of users in the source selection process, we implemented a contextualization engine and integrated it in a system o ering two services: source space exploration and distributed query processing,
Using the former, users may explore the space of sources, i.e., search and discover data sources of interest. Here, the contextualization engine fosters discovery of relevant sources during exploration. The query processing service, on the other hand, allows queries to be federated over multiple sources. See also Fig. 2 for an overview.
Interaction between both services is tight and user-driven. In particular, sources discovered during source exploration may be used for answering queries. On the other hand, sources employed for result computation may be inspected, and via contextualization other relevant sources may be found.
The data portal is based on the Information Workbench [
with statistical data/sources from Eurostat and Worldbank. This population involved an extraction of meta-data from data catalogs, represented using the VoID and DCAT vocabularies. The meta-data includes information about the accessibility of the actual data sources, which is used in a second step to load and populate the data sources locally. Every data source is stored in a triple store and accessible via a dedicated SPARQL endpoint. Overall, we have a total of more than 10000 data sources available.
Source Exploration and Selection. A typical search process starts with looking for \the right" sources. That is, a user begins with exploration of the data source space. For instance, she may issue a keyword query \gross domestic product", yielding sources with matching words in their meta-data. If this query does not lead to sources suitable for her information need, a faceted search interface or a tag-cloud may be used. For instance, she re nes her sources via entity \Germany" in a faceted search, Fig.3.
Once the user discovered a source of interest, its structure as well as entity information is shown. For example, a textual source description for GDP (current US$) is given in Fig. 4. More details about source GDP (current US$) is given via an entity and schema overview, respectively (Fig.5-a/b). Note, entities used here have been extracted by our approach, and are visualized by means of a map. Using these rich source descriptions, a user can get to know the data and data sources before issuing queries.
Further, for every source a ranked list of contextualization sources is given. For GDP (current US$), e.g., source GDP at Market Prices is recommended, Fig.5-c. This way, the user is guided from one source of interest to another. At any point, she may select a particular source for querying. Eventually, she not only knows her relevant sources, but has also gained rst insights into data schema and entities.
Processing Queries over Selected Sources. In this second component, we provide means for issuing and processing queries over multiple (previously selected) sources. Say, a user has chosen GDP (current US$) as well as its contextualization source GDP at Market Prices (Fig. 5-d). Due to her previous exploration, she knows that the former provides the German GDP from 2000 2010, while the second one features GDP from years 2011 and 2012 in Germany.
Knowing data sources that contain the desired data, the user may simply add
them to the federation by clicking on the corresponding button. The federation
can then be queried transparently, i.e., as if the data was physically integrated
in a single source. Query processing is handled by the FedX engine [
Following the example, the user may issue the SPARQL query shown in Fig. 6. Here, the user combined data from the sources GDP (current US$) and GDP at market prices. That is, while GDP data from 2000 to 2010 was retrieved from source GDP (current US$), GDP information for the years 2011 and 2012 was loaded from the data source GDP at market prices.
The results of this query can be visualized using widgets o ered by the Information Workbench. For instance, as shown in Fig.7, GDP information for Germany may be depicted as bar chart.
Future Applications of the Contextualization Engine. Besides the current usage of source contextualization, we see further applications in the future. In particular, the learned entity clusters may be used for data source search result visualization, or even for visualization of SPARQL query results. Further, SPARQL results could be ranked based on data source contextualization sources and user inputs for source selection, respectively. and schema, respectively. (c) Contextualization sources for GDP (current US$). SELECT ? year ?gdp WHERE f f g ? obs1 r d f : type qb : Observation ; wb property : i n d i c a t o r
wbi c i :NY.GDP.MKTP.CN ; sdmx dimension : r e f A r e a
wbi cc :DE ; sdmx dimension : r e f P e r i o d ? year ; sdmx measure : obsValue ?gdp .
UNION f g g ? obs2 r d f : type qb : Observation ; qb : d a t a s e t es data : tec00001 ; es property : geo es d i c : geo#DE ; sdmx dimension : timePeriod ? year ; sdmx measure : obsValue ?gdp .
FILTER(? year > "2010 01 01"^^ xsd : date ) (current US$) and GDP at Market Prices, were selected during source exploration. 5
Closest to our approach is recent work on \ nding related tables" on the Web
[
Also related are approaches on data source recommendation for source
linking, e.g., [
Another line of work is concerned with query processing over distributed
RDF data, e.g., [
Last, data integration for Web search has received much attention. Some
works target rewriting queries, e.g., [
We presented a novel approach for Web data source contextualization. For this, we adapted well-known techniques from the eld of data mining. More precisely, we provide a framework for source contextualization, to be instantiated in an application-speci c manner. By means of a real-world use-case and prototype, we show how source contextualization allows for user involvement during source selection. Based on our use-cases and data-portal system, we plan to conduct empirical experiments for validating the e ectiveness of our approach. In fact, we aim at a comparison with related work from the eld of web-table contextualization, as discussed in Sect. 5.