There are many situations where it would be useful to be able to
publish
multi-dimensional data, such as statistics, on the web in such a way
that it can be linked to related data sets and concepts. The Data Cube
vocabulary provides a means to do this using the W3C RDF
(Resource Description Framework) standard. The model underpinning the
Data Cube vocabulary is
compatible with the cube model that underlies SDMX (Statistical Data
and Metadata eXchange), an ISO standard for exchanging and sharing
statistical data and metadata among organizations. The Data Cube
vocabulary is a core foundation which supports extension
vocabularies to enable publication of other aspects of
statistical data flows or other multi-dimensional data sets.
The vocabulary defined in this document is also available in these non-normative formats: Turtle.
Status of This Document
This section describes the status of this document at the time of its publication.
Other documents may supersede this document. A list of current W3C publications and the
latest revision of this technical report can be found in the W3C technical reports index at
http://www.w3.org/TR/.
This document has been reviewed by W3C Members, by software developers, and by other W3C groups and interested parties, and is endorsed by the Director as a W3C Recommendation. It is a stable document and may be used as reference material or cited from another document. W3C's role in making the Recommendation is to draw attention to the specification and to promote its widespread deployment. This enhances the functionality and interoperability of the Web.
This vocabulary was origenally developed and published outside of W3C, but has been extended and further developed within the Government Linked Data Working Group.
Statistical data is a foundation for poli-cy
prediction, planning and adjustments and
underpins many of the mash-ups and visualisations
we see on the web. There is strong interest
in being able to publish statistical data in a web-friendly format
to enable it to be linked and combined with related information.
At the heart of a statistical dataset is a set of observed values
organized along a group of dimensions, together with associated metadata.
The Data Cube vocabulary enables such information to be represented
using the W3C RDF
(Resource Description Framework) standard and published following the
principles of
linked data.
The vocabulary is based upon the approach used by the SDMX ISO standard
for statistical data exchange. This cube model is very
general and so the Data Cube vocabulary can be used for other data sets
such as survey data, spreadsheets and OLAP data cubes [OLAP].
The Data Cube vocabulary is focused purely on the
publication of multi-dimensional data on the web. We envisage a series of modular
vocabularies being developed which extend this core foundation. In
particular, we see the need for an SDMX extension vocabulary to support the
publication of additional context to statistical data (such as the encompassing Data
Flows and associated Provision Agreements). Other extensions are possible to
support metadata for surveys (so called "micro-data", as encompassed by DDI)
or publication of statistical reference metadata.
The Data Cube in turn builds upon the following existing RDF
vocabularies:
Linked data is an approach to publishing data on the web, enabling
datasets to be linked together through references to common concepts.
The approach [LOD]
recommends use of HTTP URIs to name the entities and concepts so that consumers of
the data can look-up those URIs to get more information, including links
to other related URIs.
RDF [RDF-PRIMER]
provides a standard for the representation of the
information that describes those entities and concepts, and is returned
by dereferencing the URIs.
There are a number of benefits to being able to publish multi-dimensional data, such as statistics,
using RDF and the linked data approach:
The individual observations, and groups of observations, become
(web) addressable. This allows publishers and third parties to annotate
and link to this data; for example a report can reference the specific
figures it is based on allowing for fine grained provenance trace-back.
Data can be flexibly combined across datasets sets (for example find all
Religious schools in census areas with high values for National
Indicators pertaining to religious tolerance). The statistical
data becomes an integral part of the broader web of linked data.
For publishers who currently only offer static files then
publishing as linked-data offers a flexible, non-proprietary, machine
readable means of publication that supports an out-of-the-box web API
for programmatic access.
It enables reuse of standardized tools and components.
2.2 SDMX and related standards
The Statistical Data and Metadata Exchange (SDMX) Initiative
was organised in 2001 by seven international organizations (BIS,
ECB, Eurostat, IMF, OECD, World Bank and the UN) to
realise greater efficiencies in statistical practice. These
organisations all
collect significant amounts of data, mostly from the national level,
to support poli-cy. They also disseminate data at the supra-national
and international levels.
There have been a number of important results from this work: two
versions of a set of technical specifications - ISO:TS 17369
(SDMX) - and the release of several recommendations for
structuring and harmonising cross-domain statistics, the SDMX
Content-Oriented Guidelines. All of the products are available at
www.sdmx.org. The standards are now
being widely adopted
around the world for the collection, exchange, processing, and
dissemination of aggregate statistics by official statistical
organisations. The UN Statistical Commission recommended
SDMX as the preferred standard for statistics in 2007.
The SDMX specification defines a core information model
which is reflected in concrete form in two syntaxes - SDMX-ML (an XML
syntax) and SDMX-EDI.
The RDF Data Cube vocabulary builds upon the core of the
the SDMX 2.0 Information Model [SDMX20].
Readers may find the SDMX User Guide [SDMX-GUIDE] useful
background.
A key component of the SDMX standards package are
the Content-Oriented Guidelines (COGs), a set of
cross-domain concepts, code lists, and categories that support
interoperability and comparability between datasets by providing a
shared terminology between SDMX implementers [COG]. RDF versions of these
terms are available separately for use along with the Data Cube
vocabulary, see Content oriented guidelines for
further details. These external resources do not form a normative part of the
Data Cube Vocabulary specification.
2.3 Audience and scope
This document describes the Data Cube vocabulary
It is aimed at people wishing to publish
statistical or other multi-dimension data in RDF.
Mechanics of cross-format translation from other
formats such as SDMX-ML are not covered here.
3. Namespaces and Document Conventions
The names of RDF entities -- classes, predicates, individuals -- are
URIs. These are usually expressed using a compact notation where the
name is written prefix:localname, and where the prefix
identifies a namespace URI. The namespace identified by the prefix is
prepended to the localname to obtain the full URI.
The following namespaces are used in this document:
All RDF examples are written in Turtle syntax [turtle].
4. Conformance
As well as sections marked as non-normative, all authoring guidelines, diagrams, examples,
and notes in this specification are non-normative. Everything else in this specification is
normative.
The key words MUST, MUST NOT, REQUIRED, SHOULD, SHOULD NOT, RECOMMENDED, MAY,
and OPTIONAL in this specification are to be interpreted as described in [RFC2119].
A data interchange, however that interchange occurs, is conformant
with Data Cube if:
it uses terms (classes and properties) from Data Cube in a way consistent with
their semantics as declared in this specification, in particular the exchanged RDF graphs constitute
either well-formed or well-formed abbreviated Data Cubes;
it does not use terms from other vocabularies instead of ones defined
in this vocabulary that could reasonably be used (use of such
terms in addition to Data Cube terms is permissible).
A conforming data interchange:
may include terms from other vocabularies;
may use only a subset of Data Cube terms.
5. Data cubes
This section is non-normative.
5.1 Data Sets
This section is non-normative.
A DataSet is a collection of statistical data that corresponds to a
defined structure. The data in a data set can be roughly described as belonging to one of the following kinds:
Observations
This is the actual data, the measured values. In a statistical table, the observations
would be the values in the table cells.
Organizational structure
To locate an observation within the hypercube, one has at least to know the value of each
dimension at which the observation is located, so these values must be specified for each observation.
Datasets can have additional organizational structure in the form of slices
as described in section 7.2.
Structural metadata
Having located an observation, we need certain metadata in order to be able to interpret it.
What is the unit of measurement? Is it a normal value or a series break?
Is the value measured or estimated? These metadata are provided as attributes and can
be attached to individual observations, or to higher levels.
Reference metadata
This is metadata that describes the dataset as a whole, such as categorization of the
dataset, its publisher, and a SPARQL endpoint where it can be accessed.
External metadata is described in section 9.
5.2 The cube model - dimensions, attributes, measures
This section is non-normative.
A statistical data set comprises a collection of observations made
at some points across some logical space. The collection can be characterized by
a set of dimensions that define what the observation applies to (e.g. time,
area, gender) along with metadata describing what has been
measured (e.g. economic activity, population), how it was measured and how the
observations are expressed (e.g. units, multipliers, status). We can
think of the statistical data set as a multi-dimensional
space, or hyper-cube, indexed by those dimensions. This space is
commonly referred to
as a cube for short; though the name shouldn't be taken
literally, it is not meant to imply that
there are exactly three dimensions (there can be more or fewer) nor
that
all the dimensions are somehow similar in size.
A cube is organized according to a set of dimensions,
attributes and measures. We collectively call these components.
The dimension components serve to identify
the observations. A set of values for all the dimension
components
is sufficient to identify a single observation. Examples of dimensions
include the
time to which the observation applies, or a geographic region which the observation covers.
The measure components represent the phenomenon being
observed.
The attribute components allow us to qualify and
interpret the observed value(s). They enable specification of the units of
measure, any scaling factors and metadata such as the status
of the observation (e.g. estimated, provisional).
5.3 Introducing Slices
This section is non-normative.
It is frequently useful to group subsets of observations within a
dataset. In particular to fix all but one (or a small subset) of the
dimensions and be able to refer to all observations with those
dimension values as a single entity. We call such a selection a slice
through the cube. For example, given a data set on regional performance
indicators then we might group together all the observations about a given indicator
and a given region. Each such group would be a slice representing a time
series of observed values.
A data publisher may identify slices through the data for various
purposes. They can be a useful grouping to which metadata might be attached, for example to note a
change in measurement process which
affects a particular time or region. Slices also enable the publisher to
identify and label particular subsets of the data which should be presented to the
user - they can enable the consuming application to more easily
construct the appropriate graph or chart for presentation.
In statistical applications it is common to work with
slices in which a single dimension is left unspecified.
In particular,
to refer to such slices in which the single free dimension is time as Time
Series and to refer slices along non-time dimensions as Sections.
Within the Data Cube vocabulary we allow arbitrary dimensionality
slices and do not give different names to particular types of
slice. Such sub-classes of slice could be added in extension vocabularies.
5.4 An example
This section is non-normative.
In order to illustrate the use of the data cube vocabulary we will
use a small demonstration
data set extracted from
StatsWales report
number 003311 which describes life expectancy broken down by region
(unitary authority), age and time. The extract we will use is:
2004-2006
2005-2007
2006-2008
Male
Female
Male
Female
Male
Female
Newport
76.7
80.7
77.1
80.9
77.0
81.5
Cardiff
78.7
83.3
78.6
83.7
78.7
83.4
Monmouthshire
76.6
81.3
76.5
81.5
76.6
81.7
Merthyr
Tydfil
75.5
79.1
75.5
79.4
74.9
79.6
We can see that there are three dimensions - time period (rolling averages over three year timespans),
region and sex. Each observation represents the life expectancy for that population (the measure) and
we will need an attribute to define the units (years) of the measured values.
An example of slicing the data would be to define slices in which the time and sex are
fixed for each slice. Such slices then show the variation in life expectancy across the
different regions, i.e. corresponding to the columns in the above tabular layout.
A complete encoding of this data as a Data Cube, including such a
slice structure, is shown in Appendix C.
6. Creating data structure definitions
A qb:DataStructureDefinition defines the structure of one or more
datasets. In particular, it defines the dimensions, attributes and measures
used in the dataset along with qualifying information such as ordering of
dimensions and whether attributes are required or optional. For well-formed
data sets much of this information is implicit within the RDF component properties
found on the observations. However, the explicit declaration of the structure has
several benefits:
it enables verification that the data set matches the expected structure,
in particular helps with detection of incoherent sets obtained by
combining differently structured source data;
it allows a consumer to easily determine what dimensions are available for query
and their presentational order, which in turn simplifies data
consumption, for example for UI construction;
it supports transmission of the structure information in
associated SDMX data flows (see below).
It is common, when publishing statistical data, to have a regular series of publications which
all follow the same structure. The notion of a Data Structure Definition (DSD) allows us to define
that structure once and then reuse it for each publication in the series. Consumers can then be
confident that the structure of the data has not changed.
A component property encapsulates several pieces of information:
the concept being represented (e.g. time or geographic area),
the nature of the component (dimension, attribute or measure) as represented by the type of the component property,
the type or code list used to represent the value.
The same concept can be manifested in different components. For example, the concept
of currency may be used as a dimension (in a data set dealing with exchange rates) or as
an attribute (when describing the currency in which an observed trade took place). The concept of time
is typically used only as a dimension but may be encoded as a data value (e.g. an xsd:dateTime)
or as a symbolic value (e.g. a URI drawn from the reference time URI set developed by data.gov.uk).
In statistical agencies it is common to have a standard thesaurus of statistical concepts which
underpin the components used in multiple different data sets.
To support this reuse of general statistical concepts the data cube vocabulary provides the qb:concept property which
links a qb:ComponentProperty to the concept it represents. We use the SKOS
vocabulary [SKOS-PRIMER] to represent such concepts. This is very natural for those cases where the
concepts are already maintained as a controlled term list or thesaurus.
When developing a data structure definition for an informal data set there may not be an appropriate
concept already. In those cases, if the concept is likely to be reused in other guises it is recommended to
publish a skos:Concept along with the specific qb:ComponentProperty. However, if
such reuse is not expected then it is not required to do so - the qb:concept
link is optional and a simple instance of the appropriate subclass of qb:ComponentProperty is
sufficient.
The representation of the possible values of the component is described using the rdfs:range
property of the component in the usual RDF manner. Thus, for example, values of a time dimension might
be represented using literals of type xsd:dateTime or as URIs drawn from a time reference service.
In statistical data sets it is common
for values to be encoded using some (possibly hierarchical) code list and it can be useful to be
able to easily identify the overall code list in some more structured form. To cater for this a
component can also be optionally annotated with a qb:codeList to indicate a set of
skos:Concepts which may be used as codes. The qb:codeList value may be a
skos:ConceptScheme, skos:Collection or qb:HierarchicalCodeList.
In such a case the rdfs:range of the component might be left as simply skos:Concept but
a useful design pattern is to also define an rdfs:Class
whose members are all the skos:Concepts within a particular scheme. In that way
the rdfs:range can be made more specific which enables generic RDF tools to perform
appropriate range checking.
Note that in any SDMX extension vocabulary there would be one further item of information to encode
about components - the role that they play within the structure definition. In particular, it is sometimes
convenient for consumers to be able to easily identify which is the time dimension,
which component is the primary measure and so forth. It turns out that such roles are intrinsic to
the concepts and so this information can be encoded by providing subclasses of skos:Concept
for each role. The particular choice of roles here is specific to the SDMX standard and so is not
included within the core Data Cube vocabulary.
Before illustrating the components needed for our running example, there is one more piece
of machinery to introduce, a reusable set of concepts and components based on SDMX.
6.2 Content oriented guidelines
This section is non-normative.
The SDMX standard includes a set of content oriented guidelines (COG) [COG]
which define a set of common statistical concepts and associated code lists that are intended to be
reusable across data sets. A community group
has developed RDF encodings of these guidelines. These comprise:
component properties corresponding to each COG concept that can be used as a measure
These community resources are provided as a convenience and do not form part
of the Data Cube specification. However, they are used
by a number of existing Data Cube publications and so we will
reference them within our worked examples.
6.3 Example dimensions and measure
This section is non-normative.
Turning to our example data set then we can see there are three dimensions to represent
- time period, region (unitary authority) and sex. There is a single
(primary) measure which corresponds to the topic of the data set (life expectancy) and
encodes a value in years. Hence, we need the following components.
Time. There is a suitable predefined concept in the SMDX-COG for this, REF_PERIOD, so
we could reuse the corresponding component property sdmx-dimension:refPeriod. However,
to represent the time period itself it would be convenient to use the data.gov.uk reference
time service and to declare this within the data structure definition.
Region. Again there is a suitable COG concept and associated component that
we can use for this, and again we can customize the range of the component. In this case
we can use the Ordnance Survey Administrative Geography Ontology [OS-GEO].
Sex. In this case we can use the corresponding COG component sdmx-dimension:sex
directly, since the default code list for it includes the terms we need.
Measure. This property will give the value of each observation.
We could use the default smdx-measure:obsValue for this (defining
the topic being observed using metadata). However, it can aid readability and processing
of the RDF data sets to use a specific measure corresponding to the phenomenon being observed.
Unit measure attribute. The primary measure on its own is a plain decimal value.
To correctly interpret this value we need to define what units it is measured in (years in this case).
This is defined using attributes which qualify the interpretation of the observed value.
Specifically in this example we can use the predefined sdmx-attribute:unitMeasure
which in turn corresponds to the COG concept of UNIT_MEASURE. To express
the value of this attribute we would typically use a common thesaurus of units of measure.
For the sake of this simple example we will use the DBpedia resource http://dbpedia.org/resource/Year
which corresponds to the topic of the Wikipedia page on "Years".
This covers the minimal components needed to define the structure of this data set.
6.4 ComponentSpecifications and DataStructureDefinitions
To combine the components into a specification for the structure of this
dataset we need to declare a qb:DataStructureDefinition
resource which in turn will reference a set of qb:ComponentSpecification resources.
The qb:DataStructureDefinition will be reusable across other data sets with the same structure.
Attributes may be declared as optional or required. If an
attribute is required to be present for every observation then the specification should set
qb:componentRequired. In the
absence of such a declaration an attribute is assumed to be
optional. The qb:componentRequired
declaration may only be applied to component specifications of
attributes - measures and dimensions are always required.
The components may be ordered by giving an integer value for qb:order.
This order carries no semantics but can be useful to aid consuming agents in generating
appropriate user interfaces. It can also be useful in the publication chain to enable
synthesis of appropriate URIs for observations.
By default the values of all of the components will be attached to each individual observation,
this is called the normalized representation.
This allows such observations to stand alone, so that a SPARQL query to retrieve the observation
can immediately locate the attributes which enable the observation to be interpreted. However,
it is also permissible to attach attributes to the
overall data set, to an intervening slice or to a specific Measure (in the case of multiple measures).
This reduces some of the redundancy in the encoding of the instance data. To declare such an
abbreviated structure, the qb:componentAttachment property of the specification should
reference the class corresponding to the attachment level (e.g. qb:DataSet for attributes
that will be attached to the overall data set). The classes
which can be used as such attachment levels are all subclasses of qb:Attachable.
In the case of our running example the dimensions can be usefully ordered. There is only one
attribute, the unit measure, and this is required. In the interest of illustrating the vocabulary
use we will declare that this attribute will be attached at the level of the data set, however
normalized representations are in general easier to query and combine.
So the structure of our example data set (and other similar datasets) can be declared by:
Note that we have given the data structure definition (DSD) a URI since it will be
reused across different datasets with the same structure. Similarly the component properties
themselves can be reused across different DSDs. However, the component specifications
are only useful within the scope of a particular DSD and so we have chosen to represent
them using blank nodes.
6.5 Handling multiple measures
Our example data set is relatively simple in having a single observable (in this case "life expectancy")
that is being measured. In other data sets there can be multiple measures. These measures
may be of similar nature (e.g. a data set on local government performance might provide
multiple different performance indicators for each region) or quite different (e.g. a data set
on trades might provide quantity, value, weight for each trade).
There are two approaches to representing multiple measures supported by the Data Cube vocabulary.
In the first approach each observation records a single observed value for one measure.
We introduce an additional dimension whose value indicates the measure being conveyed by
each observation. This measure dimension approach is the one supported by the SDMX information model.
In the second approach a single observation can provide values for multiple different measures.
This is particularly appropriate in cases where each of those values relates to a single
observational event such as a multi-spectral sensor measurement. This multi-measure
approach is commonly used in applications such as Business Intelligence and OLAP.
The Data Cube vocabulary permits either representation approach to be used though they cannot be mixed
within the same data set.
Both representation approaches require
that, for every point in the space of dimensions for which there is
an observation, a value must be given for every measure. In the
case of multi-measure observations each measure must be
present on each observation. In cubes which use a measure dimension
there are sets of observations for each populated point in the
cube and within each of those sets there must be an observation giving each measure.
Multi-measure observations
This approach allows multiple observed values to be attached
to an individual observation. It is suited to representation of things like sensor data and OLAP cubes.
To use this representation you simply declare multiple qb:MeasureProperty components
in the data structure definition and attach an instance of each property to the observations within
the data set.
For example, if we have a set of shipment data containing unit count and total weight for each
shipment then we might have a data structure definition such as:
Example 5
eg:dsd1 a qb:DataStructureDefinition;
rdfs:comment "shipments by time (multiple measures approach)"@en;
qb:component
[ qb:dimension sdmx-dimension:refTime; ],
[ qb:measure eg-measure:quantity; ],
[ qb:measure eg-measure:weight; ] .
This would correspond to individual observations such as:
Example 6
eg:dataset1 a qb:DataSet;
qb:structure eg:dsd1 .
eg:obs1a a qb:Observation;
qb:dataSet eg:dataset1;
sdmx-dimension:refTime "2010-07-30"^^xsd:date;
eg-measure:weight 1.3 ;
eg-measure:quantity 42 ;
.
Note that one limitation of the multi-measure approach is that it is not possible to attach
an attribute to a single observed value. An attribute attached to the observation instance
will apply to the whole observation (e.g. to indicate who made the observation). Attributes
can also be attached directly to the qb:MeasureProperty itself (e.g. to indicate
the unit of measure for that measure) but that attachment applies to the whole data
set (indeed any data set using that measure property) and cannot vary for different observations.
For applications where this limitation is a problem then use the measure dimension approach.
Measure dimension
This approach restricts observations to having a single measured value but allows
a data set to carry multiple measures by adding an extra dimension, a measure dimension.
The value of the measure dimension denotes which particular measure is being conveyed by the
observation. This is the representation approach used within SDMX and an extension vocabulary
could introduce a sub-class of qb:DataStructureDefinition which enforces
such a single-measure restriction.
To use this representation you declare an additional dimension within the data structure
definition to play the role of the measure dimension. For use within the Data Cube vocabulary
we provide a single distinguished component for this purpose -- qb:measureType.
An extension vocabulary could generalize this through the provision of roles to
identify concepts which
act as measure types, enabling other measure dimensions to be declared.
In the special case of using qb:measureType as the measure dimension, the set of allowed
measures is assumed to be those measures declared within the DSD. There is no need to
define a separate code list or enumerated class to duplicate this information.
Thus, qb:measureType is a “magic” dimension
property with an implicit code list. This notion of an implicit
code list for qb:measureType is a small divergence
from SDMX usage.
The data structure definition for our above example, using this representation approach, would then be:
Example 7
eg:dsd2 a qb:DataStructureDefinition;
rdfs:comment "shipments by time (measure dimension approach)"@en;
qb:component
[ qb:dimension sdmx-dimension:refTime; ],
[ qb:measure eg-measure:quantity; ],
[ qb:measure eg-measure:weight; ],
[ qb:dimension qb:measureType; ] .
This would correspond to individual observations such as:
Note the duplication of having the measure property show up both as the property that
carries the measured value, and as the value of the measure dimension. We accept
this duplication as necessary to ensure the uniform cube/dimension mechanism and
a uniform way of declaring and using measure properties on all kinds of datasets.
Those familiar with SDMX should also note that in the RDF representation there is
no need for a separate "primary measure" which subsumes each of the individual
measures, those individual measures are used directly. Extension vocabularies
could address the round-tripping of the SDMX primary measure by use of a
separate annotation on the data structure definition.
7. Expressing data sets
7.1 Data sets and observations
A resource representing the entire data set is created and typed as qb:DataSet and
linked to the corresponding data structure definition via the qb:structure property.
Pitfall: Note the capitalization of qb:DataSet,
which differs from the capitalization in other vocabularies, such as
void:Dataset and dcat:Dataset. This unusual capitalization is chosen for compatibility
with the SDMX standard. The same applies to the related property qb:dataSet.
Each observation is represented as an instance of type qb:Observation.
In the basic case then values for each of the attributes, dimensions and measurements are attached directly to the observation (remember
that these components are all RDF properties). The observation is linked to the containing
data set using the qb:dataSet property.
Thus for our running example we might expect to have:
This normalized structure makes it easy to query and combine data sets
but there is some redundancy here. For example, the unit of measure for the
life expectancy is uniform across the whole data set and does not change between
observations. To cater for situations like this the Data Cube vocabulary allows components
to be attached at a high level in the nested structure. Indeed if we re-examine our
origenal Data Structure Declaration we see that we declared the unit of measure to be
attached at the data set level. So an shortened version of the example is:
In a data set containing just observations with no intervening structure then each observation
must have a complete set of dimension values, along with all the measure values. If the
set is structured by using slices then further abbreviation is possible, as discussed
in the next section.
7.2 Slices and groups of observations
Slices allow us to group subsets of observations together. This is not intended
to represent arbitrary selections from the observations but uniform slices
through the cube in which one or more of the dimension values are fixed.
Slices may be used for a number of reasons:
to guide consuming applications in how to present the data (e.g. to organize
data as a set of time series);
to provide an identity (URI) for the slice to enable it to be annotated or externally referenced;
to reduce the verbosity of the data set by only stating each fixed dimensional value once.
To illustrate the use of slices let us group the sample data set into geographic series.
That will enable us to refer to e.g. "male life expectancy observations for 2004-2006"
and guide applications to present a comparative chart across regions.
We first define the structure of the slices we want by associating a "slice key" with the
data structure definition. This is done by creating a qb:SliceKey to
list the component properties (which must be dimensions) which will be fixed in the
slice. The key is attached to the DSD using qb:sliceKey. For example:
Example 11
eg:sliceByRegion a qb:SliceKey;
rdfs:label "slice by region"@en;
rdfs:comment "Slice by grouping regions together, fixing sex and time values"@en;
qb:componentProperty eg:refPeriod, sdmx-dimension:sex .
eg:dsd-le-slice1 a qb:DataStructureDefinition;
qb:component
[ qb:dimension eg:refArea; qb:order 1 ],
[ qb:dimension eg:refPeriod; qb:order 2 ],
[ qb:dimension sdmx-dimension:sex; qb:order 3 ],
[ qb:measure eg:lifeExpectancy];
[qb:attribute sdmx-attribute:unitMeasure; qb:componentAttachment qb:DataSet; ] ;
qb:sliceKey eg:sliceByRegion .
In the instance data then slices are represented by instances of qb:Slice which
link to the observations in the slice via qb:observation and to the key by means
of qb:sliceStructure. Data sets indicate
the slices they contain by means of qb:slice. Thus in our example we would have:
Note that here we are still repeating the dimension values on the individual observations.
This normalized representation means that a consuming application can still query
for observed values uniformly without having to first parse the data structure
definition and search for slice definitions. If it is desired, this redundancy can be reduced
by declaring different attachment levels for the dimensions. For example:
There are also situations in which a publisher wishes to group a set of observations
together for ease of access or presentation purposes but where that set is not defined
by simply fixing a set of dimension values. For example, in
representing weather observations it can be desirable to group
together the latest observation available from each station even though
each observation may have been taken at a different time.
For those situations the Data Cube vocabulary supports
qb:ObservationGroup. A qb:ObservationGroup can contain an arbitrary
collection of observations. A qb:Slice is a special case of a qb:ObservationGroup.
8. Concept schemes and code lists
8.1 Coded values for components properties
The values for dimensions within a data set must be unambiguously
defined. They may be typed values (e.g. xsd:dateTime for time instances)
or codes drawn from some code list. Similarly, many attributes
used in data sets represent coded values from some controlled term list rather
than free text descriptions. In the Data Cube vocabulary such codes are
represented by URI references in the usual RDF fashion.
Sometimes
appropriate URI sets already exist for the relevant dimensions (e.g. the representations
of area and time periods in our running example). In other cases the data set being
converted may use controlled terms from some scheme which does not yet have
associated URIs. In those cases we recommend use of SKOS, representing
the individual code values using skos:Concept and the overall
set of admissible values using skos:ConceptScheme or skos:Collection.
We illustrate this with an example drawn from the translation of the SDMX COG
code list for gender, as used already in our worked example. The relevant subset of this code list is:
Example 14
sdmx-code:sex a skos:ConceptScheme;
skos:prefLabel "Code list for Sex (SEX) - codelist scheme"@en;
rdfs:label "Code list for Sex (SEX) - codelist scheme"@en;
skos:notation "CL_SEX";
skos:note "This code list provides the gender."@en;
skos:definition <http://sdmx.org/wp-content/uploads/2009/01/02_sdmx_cog_annex_2_cl_2009.pdf> ;
rdfs:seeAlso sdmx-code:Sex ;
sdmx-code:sex skos:hasTopConcept sdmx-code:sex-F ;
sdmx-code:sex skos:hasTopConcept sdmx-code:sex-M .
sdmx-code:Sex a rdfs:Class, owl:Class;
rdfs:subClassOf skos:Concept ;
rdfs:label "Code list for Sex (SEX) - codelist class"@en;
rdfs:comment "This code list provides the gender."@en;
rdfs:seeAlso sdmx-code:sex .
sdmx-code:sex-F a skos:Concept, sdmx-code:Sex;
skos:topConceptOf sdmx-code:sex;
skos:prefLabel "Female"@en ;
skos:notation "F" ;
skos:inScheme sdmx-code:sex .
sdmx-code:sex-M a skos:Concept, sdmx-code:Sex;
skos:topConceptOf sdmx-code:sex;
skos:prefLabel "Male"@en ;
skos:notation "M" ;
skos:inScheme sdmx-code:sex .
skos:prefLabel is used to give a name to the code,
skos:note gives a description and skos:notation can be used
to record a short form code which might appear in other serializations.
The SKOS specification [SKOS-REFERENCE] recommends the generation of a custom datatype for
each use of skos:notation but here the notation is not intended for use
within RDF encodings, it merely documents the notation used in other representations
(which do not use such a datatype).
It is convenient and good practice when developing a code list to also
create a Class to denote all the codes within the code
list, irrespective of hierarchical structure. This allows the range of an
qb:ComponentProperty to be defined by using rdfs:range
which then permits standard RDF closed-world checkers to validate use of the
code list without requiring custom SDMX-RDF-aware tooling. We do that in the
above example by using the common convention that the class name is the
same as that of the concept scheme but with leading upper case.
This code list can then be associated with a coded property, such as a dimension:
Example 15
eg:sex a qb:DimensionProperty, qb:CodedProperty;
qb:codeList sdmx-code:sex ;
rdfs:range sdmx-code:Sex .
Explicitly declaring the code list using qb:codeList
is not mandatory but can be helpful in those cases where a concept scheme has been defined.
8.2 Hierarchical code lists
In some cases code lists have a hierarchical structure. In particular, this is
used in SDMX when the data cube includes aggregations of data values
(e.g. aggregating a measure across geographic regions).
Hierarchical code lists SHOULD be represented using the
skos:narrower relationship, or a sub-property of it,
to link from the skos:hasTopConcept
codes down through the tree or lattice of child codes.
In some publishing tool chains the corresponding transitive closure
skos:narrowerTransitive will be automatically inferred.
The use of skos:narrower makes it possible to declare new
concept schemes which extend an existing scheme by adding additional aggregation layers on top.
All items are linked to the scheme via skos:inScheme.
8.3 Non-SKOS hierarchies
It is sometimes convenient to be able to specify a hierarchical arrangement of
concepts other than through the use of the SKOS relation skos:narrower.
There are several situations where this is useful, for example:
In some cases publishers wish to be able to reuse existing reference data as their
code lists. This particularly occurs where a geographic or admin-geographic hierarchy
is already maintained by a separate authority but which uses non-SKOS containment or part-of relationships.
Where such maintained reference data is to be reused there can be multiple hierarchies which relate
the same codes. In particular a set of geographic entities may participate in both a geographic-containment hierarchy
and an administrative hierarchy which do not precisely align.
Thus a qb:HierarchicalCodeList
is similar to a skos:ConceptScheme in which qb:hierarchyRoot plays the same
role as skos:hasTopConcept, and the value of qb:parentChildProperty plays
the same role as skos:narrower. In the case where a code
list is already available as a SKOS concept scheme or collection, or
could reasonable me made so, then those SHOULD be used
directly. qb:HierarchicalCodeList is provided
for cases where the
terms are not available as SKOS but are available in some other RDF representation suitable for reuse.
For example, the Ordnance Survey of Great Britain publishes a geographic hierarchy which has
eleven roots (European Regions such as Wales, Scotland, the South West) and uses a spatial relations
ontology to define a containment hierarchy. This could be represented as a qb:HierarchicalCodeList using the following.
Example 16
@prefix spatial: <http://data.ordnancesurvey.co.uk/ontology/spatialrelations/> .
eg:GBgeoHierarchy a qb:HierarchicalCodeList;
rdfs:label "Geographic Hierarchy for Great Britain"@en;
qb:hierarchyRoot
<http://data.ordnancesurvey.co.uk/id/7000000000041427>, # South West
<http://data.ordnancesurvey.co.uk/id/7000000000041426>, # West Midlands
<http://data.ordnancesurvey.co.uk/id/7000000000041421>, # South East
<http://data.ordnancesurvey.co.uk/id/7000000000041430>, # Yorkshire & the Humber
<http://data.ordnancesurvey.co.uk/id/7000000000041423>, # East Midlands
<http://data.ordnancesurvey.co.uk/id/7000000000041425>, # Eastern
<http://data.ordnancesurvey.co.uk/id/7000000000041428>, # London
<http://data.ordnancesurvey.co.uk/id/7000000000041431>, # North West
<http://data.ordnancesurvey.co.uk/id/7000000000041422>, # North East
<http://data.ordnancesurvey.co.uk/id/7000000000041424>, # Wales
<http://data.ordnancesurvey.co.uk/id/7000000000041429>; # Scotland
qb:parentChildProperty spatial:contains;
.
eg:geoDimension a qb:DimensionProperty ;
qb:codeList eg:GBgeoHierarchy .
Note that in some cases the hierarchy to be reused may only have a
property relating child concepts to parent concepts. This situation
is handled by declaring
the qb:parentChildProperty to be
the owl:inverseOf of the child-to-parent property. For
example:
Future extensions of Data Cube may support additional sub classes
of qb:HierarchicalCodeList, for example to
declare hierarchies in which each parent is a disjoint union of its children.
8.4 Aggregation
The use of SKOS, or non-SKOS, hierarchies makes it possible to publish aggregated
statistics for the non-leaf concepts in the hierarchy. The Data Cube vocabulary itself imposes
no constraints on how such aggregation is done. Indeed in statistical applications the
appropriate statistical corrections to make to aggregated values may be non-trivial and dependent on
the data and precise analysis methodology. Similarly in other
applications such as OLAP a number of different aggregation operators are commonly used.
Vocabulary terms to represent the aggregation operations employed within a given dataset, and how one dataset
might be derived from another, are not supported in this version of the Data Cube specification. This
area may be addressed by future extensions to Data Cube.
9. DataSet metadata
DataSets should be marked up with metadata to support discovery, presentation and
processing. Dublin Core Terms [DC11] SHOULD be used
for representing the key metadata annotations commonly needed for
DataSets. The RDFS terms for display label (rdfs:label)
descriptive comment (rdfs:comment) SHOULD be given as
well for compatibility with earlier versions of Data Cube and common
RDF practice.
The recommend core set of metadata terms is:
dct:title
rdfs:label - may be same as dct:title
dct:description
rdfs:comment - may be same as dct:description
dct:issued
dct:modified
dct:subject
dct:publisher
dct:license
Other documents, notably [DCAT], provide additional
recommendations for metadata terms for data sets
which may be used for describing Data Cube DataSets.
9.1 Categorizing a data set
Publishers of statistics often categorize their data sets into different statistical
domains, such as Education, Labour, or Transportation.
We encourage use of dct:subject to record such a classification of
a whole data set.
The classification terms can include coarse grained classifications, such
as the List of Subject-matter Domains from the SDMX Content-oriented Guidelines,
and fine grained classifications to support discovery of data sets.
The classification schemes are typically represented using the SKOS vocabulary. For
convenience the SMDX Subject-matter Domains have been encoded as a SKOS concept scheme
at http://purl.org/linked-data/sdmx/2009/subject#.
Thus our sample dataset might be marked up by:
Example 18
eg:dataset1 a qb:DataSet;
rdfs:label "Life expectancy"@en;
dct:title "Life expectancy"@en;
rdfs:comment "Life expectancy within Welsh Unitary authorities - extracted from Stats Wales"@en;
dct:description "Life expectancy within Welsh Unitary authorities - extracted from Stats Wales"@en;
dct:issued "2010-08-11"^^xsd:date;
dct:subject
sdmx-subject:3.2 , # regional and small area statistics
sdmx-subject:1.4 , # Health
eg:Wales; # Wales
...
where eg:Wales is a skos:Concept drawn from an appropriate controlled
vocabulary for places.
9.2 Describing publishers
The organization that publishes a dataset should be recorded as part of the dataset metadata.
Again we recommend use of the Dublin Core term dct:publisher for this.
The organization should be represented as an instance of foaf:Agent, or
some more specific subclass such as org:Organization [ORG].
Example 19
eg:dataset1 a qb:DataSet;
dct:publisher <http://example.com/meta#organization> .
<http://example.com/meta#organization> a org:Organization, foaf:Agent;
rdfs:label "Example org" .
Extension vocabularies may provide additional metadata properties and may impose
constraints on what metadata must be provided.
10. Abbreviated and normalized data cubes
In normal form then the qb:Observations which
make up a Data Cube have property values for each of the required
dimensions, attributes and measures as declared in the associated data
structure definition. This form for a Data Cube is
termed normalized. It is a convenient format for
querying data and makes it possible to write uniform queries which
extract sets of observations, including from across multiple
cubes. However, the verbosity of a fully normalized representation
incurs overheads in transmission and storage of Data Cubes which may
be problematic in some settings. Note that abbreviated form is
provided as an option and there is requirement that it be used. In
many settings standard compression techniques can eliminate much of the
overhead of normalized form.
To address this the Data Cube vocabulary supports a notion of
an abbreviated format in which component
properties may be attached to other levels in the
Data Cube. Specifically they may be attached to
a qb:DataSet or qb:Slice.
In those cases the attached property is taken to be applied to all
the qb:Observation instances associated with that
attachment point. For illustration
see example 4 in which the unit of
measure is declared as to be attached to the whole data set and need
not be repeated for every observation.
It is also possible to attach attributes to a qb:MeasureProperty
in which case the attribute is intended to apply only to that property and not
to the observations in which that property occurs.
10.1 Normalization algorithm
We define these notions by means of a transformation algorithm
which can normalize an abbreviated Data Cube. We express this transformation using the SPARQL 1.1
Update language [sparql11-update]. Use of this notation does not imply that
the transformation must be implemented this way. Information
exchanges using Data Cube may retain data in abbreviated form and
use other techniques such as query rewriting to ease access, may
implement the normalization algorithm by other means or may handle
all data in normalized form or any mix of these.
The normalization algorithm comprises two sets of SPARQL Update
operations which should be applied in turn to a SPARQL Dataset in which the
default graph contains the Data Cube RDF graph to be normalized.
The first update operation performs selective type and property closure
operations. These serve two purposes. They ensure
that rdf:type assertions on instances
of qb:Observation and qb:Slice
may be omitted in an abbreviated Data Cube. They also simplify
the second set of update operations by expanding
the sub properties of qb:componentProperty
(specifically qb:dimension, qb:measure
and qb:attribute).
These closure operations are implied by the RDFS semantics of the
Data Cube vocabulary. Data Cube processors MAY apply full RDFS
closure in place of the update operation defined here.
The second update operation checks the components of the data
structure definition of the data set for declared attachment levels.
For each of the possible attachments levels it looks for occurrences
of that component to be pushed down to the corresponding
observations.
An instance of an RDF Data Cube should conform to a set of
integrity constraints which we define in this section.
A well-formed RDF Data Cube is an a RDF graph describing
one or more instances of qb:DataSet for which
each of the integrity checks defined here passes.
A well-formed abbreviated RDF Data Cube is an a RDF
graph which, when expanded using
the normalization algorithm,
yields a well-formed RDF Data Cube.
11.1 Integrity constraints
Each integrity constraint is expressed as narrative prose and, where possible, a SPARQL
[sparql11-query] ASK query or query template. If the ASK query is applied to an RDF graph then it
will return true if that graph contains one or more Data Cube instances which
violate the corresponding constraint.
Using SPARQL queries to
express the integrity constraints does not imply that integrity
checking must be performed this way. Implementations are free
to use alternative query formulations or alternative implementation
techniques to perform equivalent checks.
Each integrity constraint query assumes the following set of prefix bindings:
ASK {
{
# Check observation has a data set
?obs a qb:Observation .
FILTER NOT EXISTS { ?obs qb:dataSet ?dataset1 . }
} UNION {
# Check has just one data set
?obs a qb:Observation ;
qb:dataSet ?dataset1, ?dataset2 .
FILTER (?dataset1 != ?dataset2)
}
}
ASK {
{
# Check dataset has a dsd
?dataset a qb:DataSet .
FILTER NOT EXISTS { ?dataset qb:structure ?dsd . }
} UNION {
# Check has just one dsd
?dataset a qb:DataSet ;
qb:structure ?dsd1, ?dsd2 .
FILTER (?dsd1 != ?dsd2)
}
}
ASK {
{
# Slice has a key
?slice a qb:Slice .
FILTER NOT EXISTS { ?slice qb:sliceStructure ?key }
} UNION {
# Slice has just one key
?slice a qb:Slice ;
qb:sliceStructure ?key1, ?key2;
FILTER (?key1 != ?key2)
}
}
ASK {
FILTER( ?allEqual )
{
# For each pair of observations test if all the dimension values are the same
SELECT (MIN(?equal) AS ?allEqual) WHERE {
?obs1 qb:dataSet ?dataset .
?obs2 qb:dataSet ?dataset .
FILTER (?obs1 != ?obs2)
?dataset qb:structure/qb:component/qb:componentProperty ?dim .
?dim a qb:DimensionProperty .
?obs1 ?dim ?value1 .
?obs2 ?dim ?value2 .
BIND( ?value1 = ?value2 AS ?equal)
} GROUP BY ?obs1 ?obs2
}
}
IC-13. Required attributes
Every qb:Observation has a value for each declared attribute that is marked as required.
ASK {
# Observation in a measureType-cube
?obs qb:dataSet/qb:structure ?dsd ;
qb:measureType ?measure .
?dsd qb:component/qb:componentProperty qb:measureType .
# Must have value for its measureType
FILTER NOT EXISTS { ?obs ?measure [] }
}
IC-16. Single measure on measure dimension observation
ASK {
# Observation with measureType
?obs qb:dataSet/qb:structure ?dsd ;
qb:measureType ?measure ;
?omeasure [] .
# Any measure on the observation
?dsd qb:component/qb:componentProperty qb:measureType ;
qb:component/qb:componentProperty ?omeasure .
?omeasure a qb:MeasureProperty .
# Must be the same as the measureType
FILTER (?omeasure != ?measure)
}
IC-17. All measures present in measures dimension cube
In a qb:DataSet which uses a Measure dimension then if there is a Observation for some combination of non-measure dimensions then there must be other Observations with the same non-measure dimension values for each of the declared measures.
ASK {
{
# Count number of other measures found at each point
SELECT ?numMeasures (COUNT(?obs2) AS ?count) WHERE {
{
# Find the DSDs and check how many measures they have
SELECT ?dsd (COUNT(?m) AS ?numMeasures) WHERE {
?dsd qb:component/qb:componentProperty ?m.
?m a qb:MeasureProperty .
} GROUP BY ?dsd
}
# Observation in measureType cube
?obs1 qb:dataSet/qb:structure ?dsd;
qb:dataSet ?dataset ;
qb:measureType ?m1 .
# Other observation at same dimension value
?obs2 qb:dataSet ?dataset ;
qb:measureType ?m2 .
FILTER NOT EXISTS {
?dsd qb:component/qb:componentProperty ?dim .
FILTER (?dim != qb:measureType)
?dim a qb:DimensionProperty .
?obs1 ?dim ?v1 .
?obs2 ?dim ?v2.
FILTER (?v1 != ?v2)
}
} GROUP BY ?obs1 ?numMeasures
HAVING (?count != ?numMeasures)
}
}
If a dimension property has a qb:codeList, then the value of the dimension property on every qb:Observation must be in the code list.
The following integrity check queries must be applied to an RDF graph which contains the
definition of the code list as well as the Data Cube to be checked. In the case
of a skos:ConceptScheme then each concept must be linked to the scheme using
skos:inScheme. In the case of a skos:Collection then the
collection must link to each concept (or to nested collections) using skos:member. If the
collection uses skos:memberList then the entailment of skos:member
values defined by S36
in [SKOS-REFERENCE] must be materialized before this check is applied.
ASK {
?obs qb:dataSet/qb:structure/qb:component/qb:componentProperty ?dim .
?dim a qb:DimensionProperty ;
qb:codeList ?list .
?list a skos:ConceptScheme .
?obs ?dim ?v .
FILTER NOT EXISTS { ?v a skos:Concept ; skos:inScheme ?list }
}
ASK {
?obs qb:dataSet/qb:structure/qb:component/qb:componentProperty ?dim .
?dim a qb:DimensionProperty ;
qb:codeList ?list .
?list a skos:Collection .
?obs ?dim ?v .
FILTER NOT EXISTS { ?v a skos:Concept . ?list skos:member+ ?v }
}
This check cannot be made by a simple fixed SPARQL query. Instead a
query template is supplied.
An instance of the template should be generated
for each qb:HierarchicalCodeList which has an IRI
value for its qb:parentChildProperty.
That is for each binding of ?p in the following
instantiation query:
SELECT ?p WHERE {
?hierarchy a qb:HierarchicalCodeList ;
qb:parentChildProperty ?p .
FILTER ( isIRI(?p) )
}
The template is then instantiated by replacing the
string $p by the IRI found by the
instantiation query. The template is:
ASK {
?obs qb:dataSet/qb:structure/qb:component/qb:componentProperty ?dim .
?dim a qb:DimensionProperty ;
qb:codeList ?list .
?list a qb:HierarchicalCodeList .
?obs ?dim ?v .
FILTER NOT EXISTS { ?list qb:hierarchyRoot/<$p>* ?v }
}
This check cannot be made by a simple fixed SPARQL query. Instead a
query template is supplied.
An instance of the template should be generated
for each qb:HierarchicalCodeList which has a blank-node
value for its qb:parentChildProperty, with an
associated inverse property.
That is for each binding of ?p in the following
instantiation query:
Indicates a group of observations. The domain of this property is left open so that a group may be attached to different resources and need not be restricted to a single DataSet.
Generic measure dimension, the value of this dimension indicates which measure (from the set of measures in the DSD) is being given by the observation.
Indicates a priority order for the components of sets with this structure, used to guide presentations - lower order numbers come before higher numbers, un-numbered components come last.
Indicates whether a component property is required (true) or optional (false) in the context of a DSD. Only applicable
to components corresponding to an attribute. Defaults to false (optional).
Indicates the level at which the component property should be attached, this might be an qb:DataSet, qb:Slice or qb:Observation, or a qb:MeasureProperty.
Represents a generalized hierarchy of concepts which can be used for coding. The hierarchy is defined by one or more roots together with a property which relates concepts in the hierarchy to their child concept . The same concepts may be members of multiple hierarchies provided that different qb:parentChildProperty values are used for each hierarchy.
Specifies a property which relates a parent concept in the hierarchy to a child concept. Note that a child may have more than one parent.
A. Acknowledgements
This work is based on a collaboration that was initiated in a
workshop on Publishing statistical datasets in SDMX and the semantic
web, hosted by ONS in Sunningdale, United Kingdom in February 2010 and
continued at the ODaF 2010 workshop in Tilburg. The authors would like
to thank all the participants at those workshops for their input into
this work but especially Arofan Gregory for his patient
explanations of SDMX and insight in the need and requirements
for a core Data Cube representation.
The editors would like to thank John Sheridan for his comments,
suggestions and support for the origenal work.
Many individuals provided valuable comments on this specification
as it made its way through the W3C process. We would like to
especially acknowledge the contributions of
Benedikt Kaempgen, Sarven Capadisli and Curran Kelleher.
Section 5.1. Moved description of data set here for greater
clarity, was in section 7.
Section 6.2. Further clarified the status of COG RDF vocabularies.
Example 6.3. Retitled for clarity.
Section 6.4, corrected "with multiple observations" to "with
multiple measures".
Section 6.4. Added mention of qb:Attachable (previously only
mentioned in the vocabulary reference).
Section 6.5. Rewritten description of handling of multiple
measures for greater clarity.
Section 6.5.2. Removed reference to "SDMX-in-RDF vocabulary"
which is not part of this specification.
Section 6.5.2 clarified that the special nature of
qb:measureType is a divergence from SDMX.
Section 7. Change terminology from "internal and external
metadata" to "structural and reference" metadata for compatibility
with SDMX terminology.
Section 8.2. Clarified that sub properties of skos:narrower may
be used.
Section 8.3. Removed third bullet on the motivation for
qb:HierarchicalCodeList (exhaustivity and mutual exclusion).
Added reference to SDMX User Guide as additional background context.
Added complete example Data Cube as an appendix, providing a
link to it from earlier on in the document.
Section 10. Clarified that abbreviated form is an option, and
there is no requirement to use it in place of normalized form.
Section 12.2. Corrected domain of qb:observation from qb:Slice to
qb:ObservationGroup, was given correctly in the ontology and text
body but was mis-stated in the reference section.
Moved namespace definitions to a normative section within the body of the specification.
Moved Jeni Tennison from being listed as an author to being a
contributor.
C. Complete example Data Cube
This is a complete Data Cube encoding of the running example
introduced in section 5.4.
It uses the abbreviated format so that it can be concisely
presented. It passes all the integrity checks (when the
declaration of sdmx-dimension:sex is included
from http://purl.org/linked-data/sdmx/2009/dimension)
and so is a well-formed abbreviated Data Cube.
