Data Harvesting, Curation and Fusion Model to Support Public

Service Recommendations for e-Governments

Gayane Sedrakyan

, Laurens De Vocht

, Juncal Alonso

, Marisa Escalante

, Leire Orue-Echevarria

and Erik Mannens

IMEC / IDLab, University of Ghent, Belgium

Fundacion TECNALIA Research & Innovation, ICT – European Software Institute Division, Derio, Spain

Keywords: Architectural Model, Recommendation Generation, Public Administration, Public Services, Data

Harvesting, Data Curation, Data Fusion, Linked Data, E-Government.

Abstract: This work reports on early results from CITADEL project that aims at creating an ecosystem of best

practices, tools, and recommendations to transform Public Administrations with more efficient, inclusive

and citizen-centric services. The goal of the recommendations is to support Governments to find out why

citizens stop using public services, and use this information to re-adjust provision to bring these citizens

back in. Furthermore, it will help identifying why citizens are not using a given public service (due to

affordability, accessibility, lack of knowledge, embarrassment, lack of interest, etc.) and, where appropriate,

use this information to make public services more attractive, so they start using the services. While

recommender systems can enhance experiences by providing targeted information, the entry barriers in

terms of data acquisition are very high, often limiting recommender solutions to closed systems of

user/context models. The main focus of this work is to provide an architectural model that allows harvesting

data from various sources, curating datasets that originate from a multitude of formats and fusing them into

semantically enhanced data that contain key performance indicators for the utility of e-Government

services. The output can be further processed by analytics and/or recommender engines to suggest public

service improvement needs.

1 INTRODUCTION

This work reports on early results from the

CITADEL, a H2020 European project (CITADEL

Consortium, 2017) that aims to create an ecosystem

of best practices, tools, and recommendations to

transform Public Administrations (PAs) with more

efficient, inclusive and citizen-centric services. The

CITADEL ecosystem aims to improve the processes

and policies of the PAs using what they already

know plus new data to implement what really

matters to citizens in order to shape and co-create

more efficient and inclusive public services. The

innovative ecosystem that builds on the best

practices innovates by using ICTs to find out why

citizens stop using public services, and use this

information to re-adjust provision to bring them

back in. Also, it identifies why citizens are not using

a given public service (due to affordability,

accessibility, lack of knowledge, embarrassment,

lack of interest, etc.) and, where appropriate, use this

information to make public services more attractive,

so they start using the services.

In this work we extend The DataTank (Vander

Sande, 2012), to provide the Data

Harvesting/Curation/Fusion (DHCF) component of

the platform based on which recommendations for

the utility and improvements of public services as

well as suggestions for specific services will be

generated to PAs. The DataTank provides an open

source, open data platform which not only allows

publishing datasets according to standardised

DCAT-AP guidelines and taxonomies promoted by

Open Data Support (http://opendatasupport.eu), but

also transforms the data into a variety of reusable

formats. This allows PAs to publish data in an

almost effortless manner, with maximum impact in

terms of visibility. Using this platform civil servants

will see their open datasets automatically being

crawled by other aggregation portals (a.o., the EU

open data portal) because of the DCAT-AP

compliance. The extension will include an intelligent

Sedrakyan, G., Vocht, L., Alonso, J., Escalante, M., Orue-Echevarria, L. and Mannens, E.

Data Harvesting, Curation and Fusion Model to Support Public Service Recommendations for e-Governments.

DOI: 10.5220/0006728206910698

In Proceedings of the 6th International Conference on Model-Driven Engineering and Software Development (MODELSWARD 2018), pages 691-698

ISBN: 978-989-758-283-7

691

way of harvesting and fusion of different (big) data

sources using semantics and Linked Data

technologies. In the context of CITADEL the new

DHCF component will enable the visualization and

analysis of trends for the usage of public services in

European cities, playing a key role in in terms of

suggesting improvements to the current suite of

public services.

This will allow rising the PAs’ knowledge regarding

their progress across various e-government Key

Performance Indicators (KPIs) to improve and make

more specific and evidence based on their e-

government investment plans (see KPI examples in

the section on architectural design). In long term it

would have a positive impact on more efficient and

effective e-government investment strategies of

public institutions.

While the approach that will be followed in

CITADEL for the big data analysis is not novel, the

CITADEL solution regarding big data algorithms

innovation lies on 1) the domain (public sector) in

which it will be applied, 2) the purpose for what is

created, that is, the creation of KPI reports

containing business intelligence that will be used as

input to derivate generic (semi-)automatic

recommendations to improve the processes and

policies of the PAs. The focus of this work is to

provide an architecture that will allow collecting

data from various sources in different formats (e.g.

e-Government portals, offline data, other online

sources such as social media) and fuse them into a

semantically enhanced dataset in order to facilitate

more efficient and inclusive analytics and

recommendation processes for PAs.

2 RELATED WORK

Recent R&D topics show increased interest in the

use of recommender systems for e-Government to

assist with customized suggestions for the use of

public services. While recommender systems can

enhance the user experiences by providing targeted

information, the entry barriers in terms of data

acquisition are very high (Heitmann, Hayes, 2010).

To our knowledge scientific publications describing

research-based approaches and methods for

harvesting data from multiple sources, curating and

combining different datasets as basis for

recommendations in public service domain are

largely lacking. Often, the scope of

recommendations is also limited to user models and

context variables that need to be constantly updated

by human interpreter to consider new variables and

maintain the semantics between different model

variables. To the best of our knowledge, only one

recommendation approach has been presented that

focuses on the e-Government service

recommendations that relies on semantic knowledge

using semantic ontologies. Yet, the focus of the

recommendations is limited to one specific are for

tourism (Al-Hassan et al., 2015). In this paper we

posit that harvesting of context variables and KPIs

for visualizations for e-Government service

recommendations can be extended to rely on open

data that may exist beyond such models, e.g.

anywhere from web (e.g. social media discussions)

or European portal, which may be collected and

transformed into unified dataset that is ready to be

processed by recommender engines. We posit that

linked data technologies will allow fulfilling this

task in an automated way by also maintaining the

semantics from different sources and formats.

The Semantic Web provides technologies for

knowledge representation, which can deliver Linked

Data created by multiple parties at Web scale. For

any given entity in a recommendation database, the

open world assumption means that we can harvest

more contextual information by looking up data on

the Web through link following. Because

identification of concepts happens through universal

identifiers - as opposed to local database IDs - other

parties can attach additional metadata to any concept

in order to improve recommendations.

3 ARCHITECTURAL DESIGN OF

THE PROPOSED METHOD

The KPI visualization and Report generation

component of the CITADEL ecosystem will

generate a report based on filtered KPIs. The report

will be presented as visualizations to support

recommendations to PAs. The data will be checked

for privacy sensitivity and anonymized if needed.

The process flow and possible UI mockup for some

KPI definitions/filters are shown in Figure 1 and

Figure 2 (CITADEL Consortium, 2017)

subsequently. Examples of possible KPIs include:

• KPIs to co-create: Number of users and trends

• General KPIs for improving the usage of the

current digital services in general as well as for

a specific service:

o Number of users/non users per service/per

year

AMARETTO 2018 - Special Session on domAin speciﬁc Model-based AppRoaches to vEriﬁcaTion and validaTiOn

692

o Data/information of citizens who (do not)

use the digital services (or one concrete

digital service): i.e. demographics such as

age, gender, education and computer skills,

internet access, devices used and frequencies

of using digital devices, other methods and

causes contacting governments (e.g. offline

visits, phone calls), etc.

o Built-in feedback (rating) per service/ per

type of service/ per year (e.g. problems

using services, satisfaction, etc.)

These are the possible KPIs envisioned by now. The

list of possible KPIs might be extended to include

further input from CITADEL co-creation

methodology and its subsequent component.

The Data Harvesting, Curation and Fusion

component (DHCF) in the context of KPI

visualizations will collect, store, fuse and provide

data related to the specified KPIs. The main

functionalities of this component are:

• Harvesting/loading structured file based

resources in different formats (like CSV, XLS,

JSON, XML, SQL, RDF…) but also databases

through SQL or indexes like Elasticsearch and

publishing data in different formats (like CSV,

JSON, XML, RDF…).

• Filtering resources based on their metadata, for

example based on keyword or resource type.

• Linking and fusing data sources by adding

semantic context.

• User management: set the visibility of resources

for certain resources according to user/groups.

The architecture of the DHCF component

consists of three subcomponents: The DataTank

(Vander Sande et al., 2012), RML Mapper (Dimou

et al., 2014) and RDF Store (W3C, 2014). There are

two main parts, which are able to work

independently.

Part one focuses on loading, harvesting,

managing, curating and (re)publishing structured

data sources (mainly files) in different formats. Part

two focuses on fusing the loaded/harvested sources

together.

As it can be seen in the Figure 3 there is a central

component the Resource Description Framework

(W3C, 2014) (RDF) Store that it is shared by the

two parts. The RDF Store allows storing and

querying semantically linked graphs.

Introducing semantics to structured data is important

because it allows explicitly indicating the context of

the different data sources being combined. In RDF,

The figure below gives an overview of the

components of the proposed architecture. The main

component for data collecting is the DataTank. The

DataTank allows to harvest, load and manage (or

curate) different structured data sources. It has built-

in a SPARQL (W3C, 2013) templating mechanism

to access linked data residing in an RDF Store,

which is necessary for publishing combinations and

fusions of the data sources. The main component for

fusion relies on RML Mapper (Dimou et al., 2014),

the objective of which is to convert separate sources,

that may have different structures or be in a different

format to linked data. By converting data sources to

linked data, the main component RML Mapper, is to

some extent also fusing the data or at least preparing

the data in a linked format (RDF) so that they can be

fused later, for example through a SPARQL query.

The data sources are accessed through the HTTP

interface that is provided by the DataTank. The

process followed by these components supports

three main steps:

1. Adding a data source and publishing it through

HTTP.

2. Combining data, with a SPARQL query to the

RDF Store, and adding it as a combined ‘fused’

data source.

3. Mapping one or more data sources to linked

data after they are exposed in the source

interface and loading them into the RDF store.

Figure 1: KPI visualization components in CITADEL.

Data Harvesting, Curation and Fusion Model to Support Public Service Recommendations for e-Governments

693

Figure 2: Sample UI for KPI definitions.

In addition to these steps, the component will

have two different behaviours:

• Creating new resources. There are different

ways to create new resources:

o Creating a new dataset from a single

resource using SPARQL query. The user

triggers the retrieval of the data source

through the Web UI. The DataTank will

create a reference to the data source and

expose the data source via a HTTP interface.

o Creating a new dataset from a combined

resource using SPARQL query (manual

query). The user triggers the execution of a

SPARQL query after configuring the

SPARQL query manually (currently

supported) or via selecting the properties

and original data sources to combine (to be

implemented). Like with a single data

source, the DataTank will create a reference

to the data source and expose the data source

via a HTTP interface.

o Creating a new dataset from a single

resource using SPARQL query (mapping

semantics). Mapping/linking data sources is

currently done through a mapping

document.

o Creating a new dataset from a single

resource using SPARQL query (alternative

automated solution).

The output path will follow the data naming

convention specified in the DataTank

documentation: path": "/definitions/{identifier},

where: {identifier} consists of 1 or more collection

identifiers, followed by a final resource name. (e.g.

world/demography/2013/seniors). The convention

will also support the searchability of resources.

Furthermore, the DataTank allows to categorize

resources under the following naming convention:

http://example.org/{category}/{resourcename}. The

metadata will be described in DCAT-AP format,

profiting from the functionality provided by the

DataTank.

• Retrieving resources. There is a distinction

between retrieving a single data source and a

combined data source.

o Single: A request to retrieve a single data

source is parsed through the DataTank’s

HTTP interface and results in processing the

requested data source in the requested

format and returning it to the application

where the request originated from.

o Combined: When retrieving a combined

data source, the incoming request is

translated to the configured SPARQL

Query. It is the RDFStore that contains the

linked data of all data sources that have been

mapped. So retrieving a combined data

source is only possible after mapping the

data sources that need to be combined and

specifying a SPARQL query (template)

needed for the combination of data sources.

Finally, data can as well be destroyed as required.

The harvesting and curation component will be an

internal sub-component of the KPI visualization

component. This sub-component will interact with

KPI visualization and report generation component

to receive the request of the data, with the Security

Management to request the anonymization /

encryption of certain data and with external data

sources to get the data (Figure 1).

When adding a combined resource, the DataTank

generates an RML mapping document involving the

selected data sources. The generated mapping

document maps the source files according to the

chosen and mapped properties in the data sources.

The DataTank also generates a SPARQL query that

selects the chosen output properties for this resource.

The resulting combined data source will appear in

the list of available resources. It will behave similar

to a SPARQL resource. The main difference is that

the creation of the mapping document and the

SPARQL query will be hidden from the user.

Currently the mapping documents and SPARQL

queries need to be manually configured. Figure 4

AMARETTO 2018 - Special Session on domAin speciﬁc Model-based AppRoaches to vEriﬁcaTion and validaTiOn

694

Figure 3: Harvesting/curation/fusion component architecture.

shows an example of such mapping document. The

SPARQL queries in the DataTank and the mapping

documents are on a mounted local file system or file

server. In the future, it will be possible to select data

sources as well as the properties to fuse them on and

also the target properties to map the source

properties on. It is very common that a similar

property might use a different column name to

depict the same.

Figure 4: Mapping and filtering example with manual

configuration file in JSON format.

To add a combined resource, it is necessary to

formulate a SPARQL query. RML enables reusing

mapping definitions/configurations to be used with

different formats (Figure 5).

Figure 5: Fusing different sources using RML mapping

definitions (Dimou et al., 2014).

Furthermore, RML provides a solution to model

domain-level knowledge in a scalable, integrable

and interoperable fashion by semantically

representing data derived from multiple

heterogeneous sources using the RDF framework.

RML uses uniform mapping definitions that are

independent of the references to the input data. RML

mapping definitions (Dimou et al., 2014) are 1)

reusable across different sources; 2) interoperable

across different implementations for different source

formats which allows reusing them at reduced

implementation and learning costs; 3) scalable and

extendable allowing to reference to the data extracts

and the mapping definitions in a distinct fashion by

the use of generic way of definitions for what can be

used for all possible different input sources and

scales over what cannot.

To avoid compatibility issues as well as to facilitate

versioning upgrades, the initial version of the DHCF

will rely on an interface that will link functionalities

from the DataTank and RMLMapper using REST

services. Figures 6-10 show examples of the user

Data Harvesting, Curation and Fusion Model to Support Public Service Recommendations for e-Governments

695

interfaces for the initial version of the DHCF

component. Figure 6 shows the harvesting interface

mockup that will allow easy loading of datasets from

sources such as local files, files residing on web,

database tables as queried data. Figure 7 shows the

content panel interface where the loaded resources

can be explored and altered.

Figure 6: Harvesting UI mockup.

Figure 7: Harvested resources UI mockup.

Figure 8 shows an example a search feature of the

DataTank that will be used in the DHCF initial

version. Figure 9 shows the default curation

interface which will allow adapting loaded resources

formats, e.g. converting the selected resources into

csv, xls, json, xml, … formats.

The DHCF component will have three fusion

options. By default if no semantics is defined the

resources will be fused based on header similarity.

Figure 8: Findability UI example from the DataTank.

Figure 9: Default curation UI mockup with predefined

formats using automated curation (e.g. based on header

similarity).

The second option will allow fusing files using a

mapping configuration file and/or RML mapping

definitions (Figure 4, 5).

A more advanced version will allow graphical

interface to facilitate the mapping process using

RMLMapper. An example of defining semantic

mapping is showing in Figure 10.

The unified dataset will ultimately serve an input to

KPI visualization engine to support

recommendations for the improvements of services

in general or allowing zooming into a specific

service/user context.

AMARETTO 2018 - Special Session on domAin speciﬁc Model-based AppRoaches to vEriﬁcaTion and validaTiOn

696

Figure 10: Example of a graphical interface for mapping

with RML editor (Dimou et al., 2014).

4 VALIDATION MECHANISMS

With respect to the mapping methodologies the

proposed method will benefit from the quality

assurance mechanism based on semantic linked data

technologies. In the case of the tool RML, the

quality assessment process that will be followed is

the one done for Linked Data so as data owners do

not to need to maintain and learn multiple tools.

RML has achieved that by extending RDFunit

(AKSW, «RDFUnit») which is (one of) the pioneer

tools for this job. RDFUnit (and consequently the

mapping rules that generate the Linked Data

assessment approach) mainly focuses on semantic

annotations quality assessment rather than on data

values.

With respect to data management the DHCF

component in the context of CITADEL ecosystem

will follow the principles of FAIR (FAIR Data

Management, 2016) data to enable open access,

searchability, interoperability and re-usability of the

data resources.

With respect to evaluation of the architecture for

recommendation purposes an empirical approach

will be followed to test the approach in the context

of several use cases using national, regional and

local e-Government portals described in CITADEL

project. The implemented solution will offer built-in

evaluation mechanisms considering constructs from

commonly accepted technology acceptance models

that would also allow to keep track of user

perceptions and preferences.

5 CONCLUSIONS

In this work we presented an approach to support

recommendations for the utility and improvement

needs for public service. We achieve this by

facilitating the process of collection, curation and

fusion of data originating from various sources in

different formats that may provide broader access to

KPIs relevant for public service improvement needs.

Thanks to the use of linked data technologies the

semantics between different sources and formats can

be maintained. The fused output will be ready to be

processed by analytics and visualization engines to

produce suggestions at different levels (e.g. national,

regional, local).

The approach also demonstrates a potential for the

use of personalized recommendations based on

individual profiling, for instance collecting user

variables from various sources and matching with

existing service catalogues, as well as suggesting

issues, improvements needs in the procedures, as

well as opportunities for new services, e.g. based on

collected data on user feedback, social media

activities, information on offline visits/inquiries.

Among the potential limitations of the work can be

listed the fact that the architecture does provide

immediate mechanism for addressing privacy,

ethical and legal aspects related to data collected

from sources other than open public data

repositories, which constitutes further research area.

In the more advanced version of the model this will

be covered by the privacy and security components

of the CITADEL ecosystem that will among others

deal with anonymization, data encryption/

decryption mechanisms. Yet another concern that

suggests further research is potential conflicts

between different licence policies that may arise

from the use of datasets originating from various

sources. While in principle, CITADEL fosters CC0

open data licensing scheme, use of third party data

may need further approaches to be researched. In

addition this work does not report on the specifics of

the implementation of the proposed architecture

such as scalability and performance aspects, capture

of changes in data over time, etc. While the

components and related technologies used in the

proposed architecture in principle enable these

dimensions, these specific topics remain beyond the

scope of this paper. These will however be covered

in the extended version of the work.

Another direction for future work as already

mentioned above includes empirical studies for

evaluations of the usability aspects of the component

as well as the impact of recommendations that can

Data Harvesting, Curation and Fusion Model to Support Public Service Recommendations for e-Governments

697

be achieved by exploiting the proposed architecture

in the context of a recommender system.

The more mature version of the design proposed in

this work should also consider built-in mechanisms

for capturing end-user perceptions as user

acceptance can be important to ensure the

effectiveness and continuous refinements needs and

ultimately determine its intended utility.

Furthermore, not many studies can be found in the

domain of feedback automation (Sedrakyan 2016;

Sedrakyan, Snoeck, 2016). Thus methodologies and

frameworks to extend recommendations beyond

visualization techniques by the use of automated

textual feedback targeting both facilitation of

interpretability of data visualizations as well as

procedural suggestions (Sedrakyan, Snoeck, 2017)

will constitute further research direction.

Although the focus of CITADEL project is limited

to PAs and public services, the approach can be also

inspirational beyond the domain of e-government for

the generic context of recommender systems.

ACKNOWLEDGEMENTS

This work has been supported by EC funds from

CITADEL project - Empowering Citizens To

Transform European Public Administrations

(H2020-SC6-CULT-COOP-2016-2017, EC Grant

Agreement 726755).

REFERENCES

AKSW, «RDFUnit», Available: http://aksw.org/Projects/

RDFUnit.html. Accessed 27.11.2017.

Al-Hassan, M., Lu, H., Lu, J., 2015. A semantic enhanced

hybrid recommendation approach: A case study of e-

Government tourism service recommendation system.

Decision Support Systems, 72, pp.97-109.

CITADEL Consortium, 2017. CITADEL Project. [Online]

[Accessed 04 October 2017].

CITADEL Consortium, 2017. D4.3 Initial CITADEL

Ecosystem Architecture, s.l.: s.n.

Dimou, A., Vander Sande, M., Colpaert, P., Verborgh, R.,

Mannens, E. and Van de Walle, R., 2014, April. RML:

A Generic Language for Integrated RDF Mappings of

Heterogeneous Data. In LDOW.

FAIR, European Commission Guidelines on Data

Management in Horizon 2020, Version 3.0, 2016.

Accessed 27.11.2017.

Heitmann, B., Hayes, C., 2010. Using Linked Data to

Build Open, Collaborative Recommender Systems. In

AAAI spring symposium: linked data meets artificial

intelligence (pp. 76-81).

Sedrakyan, G., 2016. Process-oriented feedback

perspectives based on feedback-enabled simulation

and learning process data analytics. PhD Thesis, KU

Leuven.

Sedrakyan, G., Snoeck, M., 2016. Enriching Model

Execution with Feedback to Support Testing of

Semantic Conformance between Models and

Requirements-Design and Evaluation of Feedback

Automation Architecture. In AMARETTO@

MODELSWARD (pp. 14-22).

Sedrakyan, G., Snoeck, M., 2017. Cognitive feedback and

behavioral feedforward automation perspectives for

modeling and validation in a learning context. In

Model-Driven Engineering and Software Development

(pp. 70-92). Springer, Cham.

Vander Sande, M., Colpaert, P., Van Deursen, D.,

Mannens, E. and Van de Walle, R., 2012. The

DataTank: an open data adapter with semantic output.

In 21st international conference on world wide web,

proceedings (p. 4).

W3C, 2014. RDF 1.1 concepts and abstract syntax.

W3C, 2013. SPARQL query language for RDF. W3C

Recommendation.

AMARETTO 2018 - Special Session on domAin speciﬁc Model-based AppRoaches to vEriﬁcaTion and validaTiOn

698