To effectively promote the exchange of scientific data, retrieval services are required to suit the needs of the research community. A large amount of research in the field of economics is based on statistical data, which is often drawn from external sources like data agencies, statistical offices or affiliated institutes. Since producing such data for a particular research question is expensive in time and money-if possible at all- research activities are often influenced by the availability of suitable data. Researchers choose or adjust their questions, so that the empirical foundation to support their results is given. As a consequence, researchers look out and poll for newly available data in all sorts of directions due to a lacking information infrastructure for this domain. This circumstance and a recent report from the High Level Expert Group on Scientific Data motivate recommendation and notification services for research data sets. In this paper, we elaborate on a case-based recommender system for statistical data, which allows for precise query specification. We discuss required similarity measures on the basis of cross-domain code lists and propose a system architecture. To address the problem of continuous polling, we elaborate on a notification service to inform researchers on newly avaible data sets based on their personal request.
At present, efforts are being made to pick up research data as bibliographic
artifacts for re-use, transparency and citation. Data publications will be
submitted to digital archives and registered in central catalogs which lays the ground
for information services to support the scientific community in finding relevant
data. Since every scientific discipline brings its own challenges in this endeavor,
specific solutions are required, so that valuable, and hence accepted, services can
be offered to the scientific community [
As a consequence, researchers look out and poll for newly available data in all sorts of directions due to a lacking information infrastructure for this domain. They exchange news on newly available data at conferences, at meetings or simply at lunch time or during coffee-break. They also revisit websites of data agencies, repositories and familiar institutes to run their personal portfolio of keyword-based queries on regular web search engine interfaces—trying to express their request for specific data sets. Although best practice at present, this strategy seems effortful and insufficient in returning a complete list of relevant data sets. This picture was shared with us in interviews we have conducted with researchers in economics.
Having catalogs of registered research data sets puts us in a good position to address the above problem and develop well-conceived search tools and services for our scientific community. Besides the fact that the catalog itself lays the ground for a more organized search, this paper tries to address the following two aspects of the identified problem: 1. Phrasing several queries with different keywords and filters of all kinds to cover the range of relevant data sets. 2. Continuous polling at regular time intervals.
The remainder of the paper is structured as follows. We review related work and decide on our approach in Section 2. Section 3 concludes the findings and formulates the functional requirements for our proposed system. Since we follow a case-based recommendation approach, we examine case base and case structure in Section 4 and elaborate on a similarity measure design on the basis of common code lists subsequently in Section 5. We propose a system architecture in Section 7. Finally, we close with conclusions and outlook in Section 8. 2
In the domain of statistical research data, one main difficulty is given by data
protection and usage rights, so that uploading entire data collections to an
independent repository causes legal problems. This is one of several reasons why we
have decided to use Semantic Web technologies for the data model, which are
strong in fine-grained referencing and in dealing with distributed data sources.
In particular, we use the RDF Data Cube Vocabulary (QB), which integrates the
SDMX standard4 and is increasingly recognized in the domain of statistics [
There are several different types of recommender systems for which a
comprehensive overview can be found at [
Another technique makes use of the items’ digital content6 which we refer to
as content-based recommendation systems [
A third system type is based on background knowledge and calculates recommendations merely on the basis of a given user query and domain-specific preference knowledge encoded in the form of rules or specifically designed similarity measures. The approach therefore does not build on usage data at all and thus is not affected by the cold-start problem. Since usage data on statistical research data sets is not easily available to us and difficult to acquire in sufficient quantity, we find this approach most suitable for our domain. The amount of statistical research data is tremendous, and the amount of usage data required scales accordingly if we plan to include all available data sets for recommendation. In addition, recommending data sets that are similar to the ones used previously may not be helpful in the scientific domain, where researchers often work on various projects simultaneously or change their research area when moving to another organization. The above described systems tend to recommend older items, because usage statistics on newer ones build up slowly8. While these drawbacks do not apply for knowledge-based recommenders, another advantage is their strength in explaining results, so that users can understand why a particular recommendation was considered relevant. Furthermore, a lot of background knowledge for statistical data is available and has even been formalized 5 urlhttp://www.amazon.com 6 be it metadata, a textual description or the digital item itself like for example in document retrieval scenarios 7 Term Frequency - Inverse Document Frequency 8 also known as the time-span problem in SDMX9, DDI10, code lists and the RDF Data Cube Vocabulary (QB), which also encourages a knowledge-based approach.
Knowledge-based recommender systems are typically constraint-based or
casebased [
Common data repositories do not yet offer recommendation features and
focus on providing full text search interfaces and filtering features. Text search
algorithms often yield scores that allow for relevance ranking and are applied on
textual fields of the respective underlying metadata model. Search criteria given
for the more structured part of the model12 are usually filtered on, meaning
that all unmatched items are removed from the ranking [
To address research objective 1, the system must provide an interface that allows
for precise specification of a data request, enabling researchers to pinpoint to
the perfect data set regardless of whether such data exists. This can be done
on the basis of the RDF Data Cube Vocabulary which provides a wide range of
predicates and attributes to formulate precise queries. The system further needs
9 Statistical Data and Metadata eXchange Language http://sdmx.org/
10 Data Documentation Initiative http://www.ddialliance.org/
11 Case-based reasoning—or case-based recommending in our case
12 e.g. creation date, size, country of origin or other domain-specific fields
13 rephrasing query terms, resetting date ranges, size parameters, geo location and
other
to know what aspects of the query are of greater, and what aspects are of lesser
significance, which can be handled with the help of user-defined weights [
The second objective can be achieved through a notification service that sends out updates on newly available data to the individual user whenever estimated relevant.
An understanding of utility must be encoded in the system, so that data not perfectly matching the user’s description can be estimated whether it yet may interest the user. In case-based recommender systems, such knowledge is encoded in similarity measures that are used to determine an estimated degree of utility of a particular case under a given query. Such measures must be designed carefully and must not make assumptions on user preferences where no foundation is given. Case-based recommenders in principle can be applied for our research objectives. However, the value of this approach depends on the question whether meaningful similarity measures can be implemented, which will be investigated in Section 5. 4
CL OBS STATUS Status of an observation with respect events such as the ones reflected in the codes composing the code list.
CL CONF STATUS Coded information about the sensitivity and confidentiality status of the data.
CL DECIMALS Gives information on the number of decimal digits used in the data.
CL FREQ Indicates the “frequency” of the data (e.g. monthly) and, thus, indirectly, also implying the type of “time reference” that could be used for identifying the data with respect time.
CL SEX Provides information on the gender.
CL TIME FORMAT Time Format as written in the SDMX-EDI and SDMX-ML messages; these codes (based on the ISO 8601 standard) indicate the type of time references used in the data. The numeric codes below (203, 102,,702) are used only in the SDMX-EDI messages; and the alphanumeric codes (P1DPT1M) only in the SDMX-ML messages.
CL UNIT MULT Unit Multiplier; indicates the magnitude in the units of measurements.
CL AREA Reference area and/or counterpart area; geographical areas, defined as areas included within the borders of a country, region, group of countries, etc.
CL CURRENCY Provides code values for currencies.
Since we use the RDF Data Cube Vocabulary to organize statistical research
data, the number of available attributes to describe research data sets is very
large, also because RDF-based descriptions are per se extensible, which might be
made use of when dealing with long tail research data of individual researchers.
Hence, we only review some of the common attributes in order to assess the
value of this approach. Table 1 gives an overview to the Cross-Domain Code
lists issued by the SDMX consortium [
This structured representation suggests to apply the local-global-principle,
which is an established paradigm in the CBR domain [
5.1
CL CURRENCY specifies the currency used in a data set. If the user explicitly queries for data sets with Euro as a currency, only such data sets should be considered suitable. Suggesting that a data set using USD would be more useful than one using CHF has no basis.14 Thus, the similarity measure for such type should be totally uninformed: sim(q, c) = 1, if q=c 0, otherwise
We find a different situtation for the area code, where groups of countries and regions build up taxonomies. Whether data about Bavaria is useful when Germany was specified in the query depends on interpretation: It may be interpreted as “data about any region in Germany is fine” or “data about Germany on the national level is needed”. Sophisticated user interfaces would be required for disambiguation, and we rather try to bypass this problem and approach a more vague but generic measure. This is supported by the consideration that even with a more precise query, the utility of a data set on other regions still remains hard to assess in general. When a data set on Bavaria is requested and a data set on Brandenburg is given, one may argue that merely the data on Bavaria represents the population a researcher wants to do research on, and any other are simply unsuitable. In contrast, one may argue as well that both regions are siblings in the sense that Germany is the subsuming parent, and a similarity value above zero appears resonable, as the data set may still reflect some of the features the researcher is after, while a data set on Idaho (USA) may not be suitable anymore, and yet another dataset on Chengdu (China) cannot be used at all. Several techniques are available for implementing a taxonomy-based similarity measure. One generic option is given to calculate a value based on the length of the shortest path. A more specific option depends on the actual query semantics and needs further consideration and discussion.
Other codes have ordered range sets. The CL DECIMALS code list denotes the number of decimals used in the data. It seems reasonable to assume that any data providing a higher number of decimals suits just as well as the data queried for, since numbers can always be rounded. In contrast, a smaller number of decimals than requested should be assumed as less suitable, as it means a lack in precision. And since the degree of precision decreases proportionally with the difference between case and query value, it suggests a typical more is better similarity measure: sim(q, c) = 1, if q ≤ c q −dc , otherwise (1) (2) where d denotes the maximal difference between query and case, which is ten in this case.
A similar case is found for the code list CL FREQ. Quarterly data can be aggregated from monthly or daily data. But if monthly data is requested, and quarterly data is given, the request is not perfectly met. Such data might yet be more useful than yearly data, so that a similar measure like the above could be reasonable. While cl decimals was based on numbers, cl freq is symbolic. 14 U.S. Dollars (USD), Swiss franc (CHF) Therefore, we could define an order and map frequency symbols to integers, so that a similar function as the above can be applied.
For the free text fields as listed in [
A so-called global similarity measure is used to aggregate the results from the attribute-level similarity calculations. As we are still in a stage of considerations, there is no point in arguing whether weighted means, Euclidean or other types of aggregation is the right method to choose. We state, however, that the measure should enable user-defined weights in the query, as it allows the researcher to emphasize on the one or other parameter.
To complete the similarity measure, we further need to specify how multiple
values are dealt with. For example, the researcher may request data on the
geographic locations France, Germany and United Kingdom. The utility of a data
set that represents the populations of England and France may then be
calculated by finding best partners for every requested country16 and build minimum,
maximum or average for the overall similarity value of the geographic attribute.
Which strategy to choose depends on the particular attribute and should be
examined carefully in evaluation with end users [
There are some special cases we need to consider. When a query specifies a value for a particular attribute, for which the case compared does not provide any value, the measure must yield some value as well. For instance, if male is specified for gender in the query, and the attribute value is not given in the case, utility should be considered zero, because the user explicitly stated that represented population should be male. It appears reasonable to take this as the default measure. However, if the user specifies free for confidentiality status, and the case does not provide any information in this regard17, the data set is not necessarily unsuitable. A reasonable way to deal with this issue could be to simply ignore this attribute in the global similarity function.
Another special situation occurs when a researcher requests data that contains values of some currency, but she does not want to specify more precisely on it. She is certain that monetary values must be part of the data while the currency unit itself is subordinate. One way to cater for this is to introduce a special value * and let sim(*,c)=1 for any case c. 15 term frequencyinverse document frequency 16 best with respect to the local similarity measure 17 due to incomplete annotation
Due to the impression that empirical research is quite data-driven, and researchers need to conintuously look out for new data sets in order to stay upto-date, we want to make some considerations on a notification service18. As our approach was to capture the researcher’s request for data in high precision, we are in a position to test incoming data sets for relevance and send out messages19. One strategy in this regard would be to notify about every data set that meets a user-defined similarity threshold. From experience, however, similarity values tend to accumulate in a particular range, which is highly dependent on the similarity measure design and the respective user query, and thus, it may be difficult to be provide a specific threshold value. In that sense, the values calculated with the help of the similarity measure should rather be regarded as scores that give means for a ranking. To bypass this problem, we suggest to send out such rankings of newly registered data sets on a regular basis as per user settings. The user may then take a closer look at the top matches and estimate their utility individually. 7
The recommender component is considered part of a larger digital archive system that manages statistical research data. Figure 2 gives an illustration of the entire system architecture, where three main components depict the relevant parts of the recommender system.
The case retrieval engine requires access to the data base that contains the
data sets and the similarity measures which should be contained in a separate
data base as to allow for independent editing whenever administrative review is
needed. The archived research data usually is maintained in its specific data
format, which in our case is based on RDF20. If the retrieval engine is implemented
using sequential similarity calculation, the data repository can be accessed as a
case base directly, since no further indexing is required. This, however, leads to
long computation times in case of a large case base. For more efficient retrieval,
more optimized methods like Case Retrieval Nets [
The notification service needs access to the users’ notification queries and their e-mail addresses, which are stored in the user preferences data base. The component should be notified whenever updates occur on the data repository, so that new data sets can be tested for relevance immediately. 18 cf. Google Alerts 19 whenever a relevant candidate is detected or collated, per e-mail, twitter or other channel 20 Resource Description Framework
A detailled discussion on the user interface exceeds the scope of this paper.
However, it must provide for query specification, display of results and
configuration of user settings regarding the notification service as discussed in Section 6.
Furthermore, we suggest to integrate explanation features in order to provide
transparency to the user on how results were retrieved. A generic interface
design and implementation can be found at [
We have examined some of the common code lists for statistical data with respect
to their specification and found indicators that motivate a particular similarity
measure design. For some of them we were able to reason a specific design,
whereas other code lists are difficult to make assertions on and suggest rather
uninformed measures. Eventually, a final assessment on utility of a particular
data set can only be done by the researcher. A similarity measure can only
approximate a common sense of utility [
Another common practice in empirical research is the use of proxy variables, where some data highly correlates with other. Such information could be useful for recommending relevant data. A similarity measure could again be extended to make use of such relations if represented in the data model.
The proposed recommender system is based on the RDF Data Cube Vocabulary. The user is therefore in a position to specify precisely on the kind of data needed, and the system has the required means to assess suitability of available data sets. In addition, provided the measure reflects a reasonable understanding of utility, the introduced notification service helps researchers keep up to date and thus, both research goals defined in Listing 1 were met. Nevertheless, an evaluation is yet to be carried out, which is subject of future work. With further progress on a research data management infrastructure and the continuing exchange with the scientific community, we will get a clearer picture on the applicability of this approach.
Eventually, a prototype is needed in order to gain feedback from the research community we are addressing, which we consider implementing as we proceed with the reasearch on a data management infrastructure.