Many national and international organizations today leverage semantic web technologies to make statistical datasets available as Linked Open Data (LOD). A key advantages of this approach is that the data not only becomes publicly available, but also machine-readable and hence suitable for automated discovery and exploration. Whereas this has great potential to support interesting use cases, it remains difcult for end users today to utilize and combine these statistical Linked Data. Three challenges are: (i) directing users to relevant data sources based on a speci ed location; (ii) facilitating data integration despite a lack of outgoing links between datasets; and (iii) o ering exible means to integrate and aggregate data from various sources. As time and location are highly relevant dimensions in most statistical data, we address the identi ed challenges by rst constructing geographical metadata for statistical sources. Following a mashup approach, we introduce mechanisms to recommend interesting datasets to end users and automatically enable data integration, visualization, and comparisons based on userde ned criteria.
In recent years, statistical data communities have increasingly adopted
semantic web technologies in order to formalize and link their published datasets. In
particular, many organizations make use of the W3C RDF Data Cube
vocabulary [
(i) There are a large and growing number of statistical sources covering a wide range of domains. Those are valuable assets, but without an appropriate catalogue, users are unable to discover relevant statistical information.
(ii) Each statistical LOD dataset typically uses a distinct terminology and does not consistently include outgoing links to other sources. Consolidating equivalent entities stored in di erent sources therefore poses a signi cant challenge. For example, Ireland's 2011 census data1 includes thousands of geographical entities (county, province, town, city, etc.), none of which are linked to any external resources. Other LOD sources refer to Irish cities with di erent URIs. Collecting all statistical information on a city available as LOD is therefore difcult.
(iii) Current statistical data visualization applications typically focus on displaying a given dataset of a particular granularity (e.g. country level, city level, etc.). They do not, however, provide mechanisms for locating relevant datasets suitable for meaningful comparisons.
This paper addresses these issues, focusing on the geographical dimension which is highly relevant in most statistical datasets. In particular, we use geographical points as input for the data source discovery process. Based on a given location, we can nd appropriate datasets that provide statistical information. For instance, a user may locate his/her home on a map and trigger a retrieval process that yields datasets that contain data about the speci ed location. Using the supplied location, we determine the geographical entities that the location is situated in at di erent levels and allow users to compare data on the respective areas to appropriate similar areas, e.g., compare the population of their country to that of others, or compare income levels in their district to that in other districts in the same city.
To deal with the rst issue of discovering relevant statistical datasets, we analyze available SPARQL endpoints to construct geographical metadata, which connects each dataset to the related locations. We use Google Geocoding API2 to collect information of a geographical area based on their name.
The second issue, i.e., lacking support for data integration due to missing or inconsistent links, is addressed by using our constructed data catalogue. This catalogue uses a consistent unique identi er for the same location, which can refer to di erent identi ers in disparate datasets. This approach is possible even in cases where the original location entities do not reference common external entities.
To allow users to automatically integrate data from di erent statistical sources
and overcome the third issue, we make use of the Linked Widget platform [
The remainder of this paper is organized as follows. In Section 2, we discuss the development of a geographical metadata catalogue. Section 3 illustrates our mashup approach with a collection of widgets and three practical use cases. Section 4 provides pointers to related work and we conclude in Section 5 with an outlook on future research. 1 http://data.cso.ie/datasets/index.html 2 https://developers.google.com/maps/documentation/geocoding/
Countries are typically subdivided into hierarchical administrative units, e.g., regions, provinces, cities, districts, communes. Current statistical LOD sources usually provide information describing speci c characteristics of these administrative units in chronological order. Time and location are therefore typically the most important dimensions for this kind of statistical data.
We therefore develop a catalogue for spatial statistical data. If users provide an address, e.g., Donaufelder Strasse 54, Austria, or simply de ne a point on a map, we detect corresponding administrative areas and use those to select relevant datasets. We then provide suggestions to compare these areas with other areas at the same administrative level. For example, users can compare the population of their district to that of other districts in their city, or contrast the income of their city to other cities in the country ; on a higher level, they can compare the happiness index of their country to that of other countries in the world.
We use SPARQL endpoints as input of the catalogue creation process and perform four steps: (i) identify all datasets of the source; (ii) identify all dimensions and measures for each dataset; (iii) from the set of detected dimensions, determine the location dimension, (iv) for each text value of the dimension location, identify its administrative area and create a unique resource in the catalogue. We follow the results returned by the Google Geocoding API to deal with di erent geographical hierarchies of countries, ensuring consistent entity classi cation.
The metadata structure of a dataset is illustrated in Listing 1. A de nitive mapping between the administrative area and the resource URI in the metadata catalogue must exist so that same locations from di erent sources have the same unique URI in the metadata catalogue. This facilitates integration and aggregation. We de ne the mapping rule and provide a URI mapping service to convert an administrative area to the corresponding URI in the catalogue. PREFIX qb : <http :// purl . org / linked - data / cube # > PREFIX lc : <http :// location - based - catalogue . ifs . tuwien . ac . at /> [] 3 3.1 a lc : SPARQLEndpoint ; lc : hasDataSet [ lc : describes qb : dimension qb : measure qb : attribute [ a [ a [ a [ a lc : AdministrativeArea ]; qb : DimensionProperty ]; qb : MeasureProperty ]; qb : AttributeProperty ]; ]
Listing 1: A SPARQL query for terminal matching
To enable users to exibly select and combine statistical datasets and synthesize desired information, we follow a visual programming paradigm implemented in the Linked Widgets platform { a widget-based mashup platform. Its key elements are Linked Widgets, an extension of standard web widgets backed by a semantic model that follows Linked Data principles. This model describes data input/output and metadata (such as data provenance and licensing terms) that is useful for widget search and auto composition features.
There are three types of widgets, i.e., data, process, and visualization widgets. The platform provides a graphical interface for creating a data ow and composing various applications by connecting widgets in di erent ways. Other stakeholders can develop widgets independently and contribute widgets to the platform to extend its functionality. Whereas existing similar platforms are oriented more towards low-level data processing with generic widgets, the platform is more problem-oriented and supports modeling on a higher, semantic level. 3.2
In the platform, widgets are grouped into widget collections. Each collection addresses a di erent problem domain. We developed a collection3 for statistical data exploration based on spatial context. Each widget can have multiple inputs, but only a single output. The output of a widget can serve as input of another one, if their semantic models can be matched, i.e., they have a certain level of overlap. Our exemplary statistical widget collection consists of the following four widgets:
Spatial Entity Recognizer (data widget): This widget accepts an address text or a user-de ned location point as an input and calls the Google Geocoding API service to obtain the corresponding spatial entities at di erent levels, i.e., country level, administrative area level one to ve, and sublocality level. It then uses the URI mapping service to return the URIs of these areas from the catalogue. Note that these URIs may not be in the catalogue, because the catalogue does not create instances of all locations in the world; it only contains instances of locations that exist in the analyzed statistical datasets.
Spatial Data Locator (process widget): This widget returns a list of datasets related to the input spatial entity. It contains options to lter the output datasets based on di erent domains, e.g., census, transportation, or income.
Spatial Comparator (process widget): From the input of two administrative areas, this widget returns a list of datasets that contain information on the two areas that can be compared.
Google Chart (visualization widget): This widget presents the input of a single dataset or a list of datasets in appropriate charts.
All datasets returned by the two process widgets follow the W3C RDF Data Cube vocabulary and represent the data in JSON-LD format. 3 http://linkedwidgets.org/MashupPlatform.html?widgetCollectionId= SpatialStatisticalCollection
Spatial
Entity Recognizer
Spatial Comparator
Google Chart Sample mashups created from the four widgets are shown in Fig. 1: (i) discovery of statistical information on an area (at di erent administrative levels) based on a user's address;4 (ii) users select di erent criteria to compare a pair of areas, which have the same administrative level and the same parent area.5 4
In the research eld of Geographic Information Systems (GIS) the problem of
spatial dataset discovery and integration has also been addressed. Hariharan et
al. [
In the Linked Data domain, with the growing number of statistical SPARQL
endpoints available, several useful applications have emerged from the semantic
web community. The Linked Data Query Wizard is a web-based tool for
displaying, accessing, ltering, exploring, and navigating Linked Data [
Zapilko et al. [
Similar to the related work, we try to provide an easy to use interface for lay users to discover and explore statistical datasets. However, in contrast to current implementations, we want to provide the user with greater control to in uence the process of dataset discovery, on the one hand, and data exploration, on the other hand. This is achieved by employing an approach based on widgets. Users can apply widgets to control the data processing ow from the beginning, i.e., dataset discovery, to the end, i.e., data visualization, allowing for the creation of many di erent scenarios. Furthermore, through creating a metadata catalogue, we o er a exible means to compare di erent datasets, which are related via administrative levels, or statistical criteria they o er. 5
This paper addresses two crucial issues in statistical Linked Data: (i) Where and how can users discover statistical data? (ii) How can data reconciliation, ontology matching and instance matching be performed with statistical data?
We address these issues by building a data catalogue, describing identi ed datasets, and nally fostering data discovery by adopting a mashup-based approach. Our approach is built around geographical context information that users provide by specifying a geographical point or an address. In future work, we will focus on the temporal dimension of the data. For instance, users provide a certain timespan they are interested in and our system suggests datasets, which support this timespan. Furthermore, we will support users in discovering datasets based on keyword search.
The current mashups work well with datasets that have a fairly limited number of dimensions. For larger datasets, it is di cult to compare di erent areas in a single chart. We can standardize locations through available geocoding APIs, but to standardize dimensions or measures that do not have the same meaning is still an open challenge. For instance, the year dimension has di erent URIs in di erent statistical data sources. At present, our implementation is applicable for datasets of the same statistical LOD source only and we need to improve this in future.
We currently develop a Dataset Recommender widget which accepts a single spatial entity as input and nds all spatial entities that share the same parent. Then, based on the criteria chosen by the user, it returns a list of datasets that contain statistical information that relate to the areas. When connecting its input with the output of Spatial Entity Recognizer and its output with the input of Google Chart, we have another mashup enabling users to compare their area to appropriate, automatically identi ed, other areas.
Additionally, we plan to automatically create outer links to DBpedia for the location entities of the metadata catalogue. This would allow other developers to more easily integrate statistical data with the Web of Data.
Finally, it will be necessary to evaluate precision and recall of the location detection process during the metadata construction process in a dedicated experiment.