This is the authors’ preprint version. The article is accepted for publication in the Special Issue on Data Transparency in the ACM Journal of Data and Information Quality (JDIQ). ©2021 arXiv:2105.09312v2 [cs.DB] 21 May 2021 Knowledge-driven Data Ecosystems Towards Data Transparency SANDRA GEISLER, Fraunhofer FIT and RWTH Aachen University, Germany MARIA-ESTHER VIDAL, TIB-Leibniz Information Centre for Science and Technology, Gerrmany CINZIA CAPPIELLO, Politecnico di Milano, Italy BERNADETTE FARIAS LÓSCIO, Federal University of Pernambuco, Brazil AVIGDOR GAL, Technion Israel Institute of Technology, Israel MATTHIAS JARKE, RWTH Aachen University and Fraunhofer FIT, Germany MAURIZIO LENZERINI, Sapienza Università di Roma, Italy PAOLO MISSIER, Newcastle University, United Kingdom BORIS OTTO, TU Dortmund University, Germany and Fraunhofer ISST, Germany ELDA PAJA, IT University of Copenhagen, Denmark BARBARA PERNICI, Politecnico di Milano, Italy JAKOB REHOF, TU Dortmund University, Germany and Fraunhofer ISST, Germany A Data Ecosystem offers a keystone-player or alliance-driven infrastructure that enables the interaction of different stakeholders and the resolution of interoperability issues among shared data. However, despite years of research in data governance and management, trustability is still affected by the absence of transparent and traceable data-driven pipelines. In this work, we focus on requirements and challenges that data ecosystems face when ensuring data transparency. Requirements are derived from the data and organizational management, as well as from broader legal and ethical considerations. We propose a novel knowledge-driven data ecosystem architecture, providing the pillars for satisfying the analyzed requirements. We illustrate the potential of our proposal in a real-world scenario. Lastly, we discuss and rate the potential of the proposed architecture in the fulfillment of these requirements. Additional Key Words and Phrases: Data transparency, data ecosystems, data quality, trustability 1 INTRODUCTION Industrial digitalization and the use of information technologies in public and private sectors provide evidence of the pivotal role of data. However, despite the paramount relevance of datadriven technologies, organizations demand alliance-driven infrastructures capable of supporting controlled data exchange across diverse stakeholders and transparent data management. Data Ecosystems (DEs) are distributed, open, and adaptive information systems with the characteristics of being self-organizing, scalable, and sustainable [27]. While centered on data, the main Authors’ addresses: Sandra Geisler, sandra.geisler@fit.fraunhofer.de, Fraunhofer FIT and RWTH Aachen University, Germany, Schloss Birlinghoven, Sankt Augustin, 53757; Maria-Esther Vidal, maria.vidal@tib.eu, TIB-Leibniz Information Centre for Science and Technology, Gerrmany, Welfengarten 1B, Hannover, 30167; Cinzia Cappiello, cinzia.cappiello@polimi.it, Politecnico di Milano, Italy, piazza Leonardo da Vinci 32, Milano, 20133; Bernadette Farias Lóscio, bfl@cin.ufpe.br, Federal University of Pernambuco, Brazil, Cidade Universitaria, Recife/PE, 50740-560; Avigdor Gal, avigal@ie.technion.ac.il, Technion Israel Institute of Technology, Israel, Technion City, Haifa, 32000; Matthias Jarke, jarke@dbis.rwth-aachen.de, RWTH Aachen University and Fraunhofer FIT, Germany, Ahornstrasse 55, Aachen, 52056; Maurizio Lenzerini, lenzerini@diag.uniroma1.it, Sapienza Università di Roma, Italy, via Ariosto 25, Roma, I-00185; Paolo Missier, paolo.missier@ncl.ac.uk, Newcastle University, United Kingdom, Firebrick Avenue, Newcastle upon Tyne, NE4 5TG; Boris Otto, boris.otto@cs.tu-dortmund.de, TU Dortmund University, Germany, Otto-Hahn-Str. 12, Dortmund, 44227, Fraunhofer ISST, Germany, Emil-Figge-Straße 91, Dortmund, 44227; Elda Paja, elpa@itu.dk, IT University of Copenhagen, Denmark, Rued Langgaards Vej 7, Copenhagen S, DK-2300; Barbara Pernici, barbara.pernici@polimi.it, Politecnico di Milano, Italy, piazza Leonardo da Vinci 32, Milano, 20133; Jakob Rehof, jakob.rehof@cs.tu-dortmund.de, TU Dortmund University, Germany, Otto-Hahn-Str. 12, Dortmund, 44227, Fraunhofer ISST, Germany, Emil-Figge-Straße 91, Dortmund, 44227. 111 111:2 S. Geisler, M.-E. Vidal et al. concern with DEs is about knowledge generation and sharing. Thus, they aim to solve issues like learning from unstructured and heterogeneous data, and construct new abstractions and mappings. They may also offer various data-centric services, including query processing and data analytics. DEs are equipped with computational methods to exchange and integrate data while preserving personal data privacy, data security, and organizational data sovereignty. The report of the Dagstuhl Seminar 19391 (September 22-27, 2019)1 on Data Ecosystems: Sovereign Data Exchange among Organizations [6] contains summaries of discussions and abstracts of talks from the seminar on various topics, including requirements, use cases, and architectures. Diverse reference architectures rely on DE foundations [2, 26]. Keystone player-driven data ecosystems and B2C platforms like Google, Alibaba, or Facebook are hugely successful. In contrast, the adoption of alliance-driven platforms which aim at more equitable control and data sharing [28] is still lagging, even in crucial domains such as data-driven B2B engineering collaboration [19] or biomedicine. This paper focuses on the alliance-driven setting, even though many addressed issues occur also in the other category. A few works address general data quality (DQ) aspects of DEs (e.g., [4, 12, 23, 35]). In [12], the authors focus on open DEs and claim that data availability and DQ need to be guaranteed, so as to prevent users to be hesitant to use data. Kitsios et al. [23] depict DQ assessment as one of the fundamental components for building and maintaining a DE. DQ assessment requires the definition of a DQ model composed of DQ dimensions and metrics. Several DQ dimensions have been defined in the literature, as discussed in [4]. Many dimensions have a possible impact on data fairness and trustability, in particular completeness, accuracy, and consistency, which have a significant impact both, on transparently processing data for analysis and on data pipelines. The lack of accountability for data transparency is one of the severe limitations of existing interoperable methods and represents a critical aspect of data quality. This paper starts from the hypothesis that these limitations could be a significant reason for the slow adoption of DEs. To account for data transparency, the rest of this paper offers the following contributions: (1) an analysis of the specific requirements arising for DEs and from DEs regarding data governance and transparency aspects; (2) a new form of networks of knowledge-driven DEs towards trustworthiness and wider adoption; (3) challenges that stem from the identified requirements; and (4) an assessment of how the various DE types address these requirements. 2 REQUIREMENTS OF TRANSPARENT DATA ECOSYSTEMS MOTIVATED BY AN EXAMPLE We motivate the need for expressive data ecosystems with an example from the health domain. Subsequently, we grasp the requirements demanded for data transparency in similar scenarios. 2.1 Motivating Example Consider a use case of multi-site clinical studies as an example to illustrate the impact that managing multiple stakeholders have on interoperability and transparency requirements. In these studies, several parties are involved; they include clinics, resident doctors, data scientists, patients, study nurses, quality assurance, researchers, and care services. A stakeholder may have multiple sources generating data. For example, clinicians conduct examinations and collect, amongst others, sensor readings, medical images, test results, and diagnostic reports. These data collections are processed (e.g., transformed, curated, and integrated); for transparency reasons, they are potentially annotated with meta-data, domain vocabularies, and data quality values. Data is analyzed to uncover insights that can support clinicians to conduct thoughtful diagnostics and effective treatments. Data management tasks are also influenced by the organization’s 1 http://www.dagstuhl.de/19391 Knowledge-driven Data Ecosystems Towards Data Transparency 111:3 regulations or higher instances, such as regulations for data protection or rules defined by the hospital, and strategic decisions. Patients may require transparency about both, their treatment and the privacy protection of their data in cross-clinical studies. Each of these data management tasks brings up already multiple challenges for data transparency and data quality management. Additionally, data is exchanged between stakeholders to fulfill the goals of the studies; data collected from the various sites has to be pseudonymized and integrated to be audited by quality assurance. But transparency of these processes must be maintained to protect against scientific fraud. Further integration with data from additional parties, such as health insurance companies, may be needed to be finally analyzed by study researchers. The study setting corresponds to a network of knowledge-driven DEs. This network aligns the stakeholder DEs and their data; it also uses meta-data to describe the network and its constituents. Furthermore, the network is influenced by regulations, contracts, or agreements specific to the study at hand. They may include participation agreements created by insurance companies and patients consent forms authorizing data usage for specific studies. Heterogeneity issues across the different network DEs impose challenges for DQ management. Moreover, documenting computational methods performed to assess and curate data quality issues is crucial to guarantee data transparency. 2.2 Requirements Analysis The motivating example highlights the multiple issues that a DE needs to cover in order to enhance trustability of the involved stakeholders. These issues are not only present in biomedical applications, but rather exist in any application where crucial decisions are driven by data [3]. Based on literature and reports from current European data sharing and data space projects [3, 10, 29], requirements can be categorized along data management, organizational aspects, and legal and ethical issues. In terms of data management, tackling the challenges outlined in the motivating example, demands (meta)data sharing among different stakeholders in a secured and traceable manner. At an organization level, trustable data exchange requires complex ecosystems that underlie organizational-specific business models, processes, and strategies to enforce sovereignty, privacy, and protection of both, data and analytical outcomes. Furthermore, sharing sensitive and personal data, e.g., clinical records, should comply with data protection regulations and legal compliance at national and international levels. More importantly, accounting for ethical decisions made by stakeholders and algorithms is crucial and the ability to provide reliable and verifiable explanations of these decisions. Meeting these requirements at a legal and ethical level empowers DEs to safeguard data privacy and mitigate unfairness in data-driven pipelines. Moreover, the satisfaction of these requirements provides the foundations for ensuring that clinical data is only used according to consents given by these data owners, i.e., the patients. Next, each of these three requirement categories is described in more detail. Data management requirements. DEs, as described in the motivating example, demand sharing of data with a high variety (e.g., in terms of type, structure, size, or frequency). The requirements listed in this category concern both, data and metadata. Data quality management has to be able to (DMR1) handle all kinds of data and offer common DQ tools to describe, query, and assess quality values for the data. In the medical domain for example, this comprises unstructured data, such as images and texts, but also highly structured data from databases, csv files, and data streams. Additionally, (DMR2) the data has to be fit for sharing. Data has to exhibit quality values, which fulfill a certain quality standard, suitable for sharing it in a defined context. Data consumers, especially in data markets, have thereby the possibility to query data based on its quality. Hence, data can be rejected, if it does not satisfy the negotiated standards. In the motivating example, this could mean that the reading center rejects the data, because important values are missing, 111:4 S. Geisler, M.-E. Vidal et al. i.e., the completeness of the data set is too low. Furthermore, data transparency plays a crucial role for enhancing trust for all stakeholders. (DMR3) Data transparency has to be enabled from the origin of the data until its usage. At any time in a data-driven pipeline, the current meaning of the data has to be available, as well as metadata describing data transformations made by the different components of the pipeline. This explicitly includes traceability and transparency of algorithms and their results (e.g., for data curation and integration, or for prediction). Consider for example the use of data from cancer registries by researchers and other registries. Both need to know explicitly how the data has been acquired and modified to estimate its value for the research at hand. Potential conflicts between data transparency and company secrets or privacy may exist. Thus, transparency must be offered to all the stakeholders according to their role in the DE, and in terms of consent management and usage control. Hence, (DMR4) data quality management needs to take trade-offs into account and provide dimensions and assessment metrics that enable stakeholders to rate the possible impact of, e.g., data curation. Anonymized medical data for example may loose its value for further research if important attributes are eliminated from a data set. Lastly, data integration and querying over multiple data sources and across organizations are required in a multitude of scenarios. For this, mappings among data sources are defined either manually or (semi-)automatically by schema matching. For data quality management this implies several aspects, but basically (DMR5) stakeholders should be part of the loop of data quality assessment. They should be able to rate the quality of every step in a data-driven pipeline, e.g., schema and entity mappings or query answers. The automatic matching between huge medical taxonomies, e.g., for decision support systems, may be very error-prone as the taxonomies per se have quality problems. (DMR6) The impact of adding a new component to a DE should be measurable. The DEs and their stakeholders should be able to rate the impact of, e.g., the information gain of adding a new data set. This is a crucial aspect especially when considering to pay for a costly data set or when the integration of the data set requires a lot of upfront effort in terms of data cleaning, transformation, or data integration. Organizational-centric requirements. In cases of sensitive data exchange and processing, data must be transparently used according to organizations’ policies, as well as its business models and strategies. (OCR1) Enabling data governance is crucial for the appropriate data exchange and sharing according to the organizations’ strategies and business models. (OCR2) Ensuring traceability of data sovereignty is essential to increase trust among stakeholders. Again this has to be ensured throughout the whole data processing pipeline including data quality assessment and curation. For example, the willingness of patients to use applications or participate in studies may be increased by giving them the opportunity to enforce access and usage policies. Furthermore, (OCR3) business, certification, and utility models need to be established to certify, based on data quality values and other characteristics, the monetary value of exchanged and transformed data. The monetary value of medical data is manifold, e.g., data from clinical studies may be of interest to other parties, such as pharmaceutical or insurance companies, to create or improve products. (OCR4) Adherence to data and data processing standards is required to enhance DQ and interoperability across stakeholders, which is specifically important in the medical domain. Standards such as FHIR2 have made an important step forward reaching these goals in the clinical domain. (OCR5) Flexible DQ management for different coordination and negotiation models among stakeholders (e.g., clinics, data scientists, and insurance companies), and considering the evolution of these models over time. Legal and ethical requirements. As shown in the motivating example, respecting personal data privacy and security during data management, exchange, and analytics impairs requirements 2 http://www.hl7.org/fhir/index.html Knowledge-driven Data Ecosystems Towards Data Transparency 111:5 at both, legal and ethical levels. Both categories of requirements are aligned with the European Union guidelines for Trustworthy AI [14]. (L&ER1) Providing expressive legal frameworks for exchanged data, including legal references, responsibilities, licenses, and ethical guidelines is essential. (L&ER2) Accounting and mitigating bias and fairness ensure that the outcomes of the execution of the system components are independent of sensitive attributes (e.g., gender, age, ethnicity, or health conditions) and augment confidence in the impartiality of the system behavior. (L&ER3) Endeavouring safeness and robustness of the decisions made by each component that exchanges, processes, or analyzes data. Thus, the deployment of data-driven pipelines and their outcomes will guarantee the compliance with ethical guidelines (e.g., the respect of the patient consents), and the misuse reduction that could conduce to data quality issues, data privacy violation, and unfairness. (L&ER4) Enforcing data protection and ownership safeguards privacy, sovereignty, and legal compliance with licenses and regulations for data sharing, exchange, and processing. Thus, the satisfaction of this requirement will ensure that clinical data is used by the distinct parties as indicated in the patient’s consents. (L&ER5) Pursuing diversity and non-discrimination in data collections shared, exchanged, and processed by data-driven pipelines. As a result, the risk of excluding specific entities (e.g., patients with a given health condition) is mitigated, and the chances of covering all the representative entities of the population increase. (L&ER6) Trackability of regulations compliance in the way that each data-driven decision can be documented and validated in terms of legal regulations, business models, and ethical guidelines. Lastly, (L&ER7) Trustworthiness and Reliability of data-driven pipelines demand the accurate measurement, validation, and interpretation of each of the decisions taken by the system components in compliance with legal and ethical guidelines of the stakeholders. Thus, data owners (e.g., patients, insurance companies, and researchers) will be able to trace management and analysis methods performed over their data. As illustrated in the next section, this paper positions networks of knowledge-driven DEs as alliancedriven decentralized infrastructures empowered with components to satisfy the requirements listed above. Lastly, section 4.1 presents the challenges to be achieved to meet these requirements. 3 KNOWLEDGE-DRIVEN DATA ECOSYSTEMS: OVERVIEW & ARCHITECTURE The literature defines DEs in different ways. For instance, Oliveira and Farias Lóscio [27] define Data Ecosystems as a łset of networks composed of autonomous actors, which consume, produce, or provide data or other related resources.ž Other definitions add that the results created by the consumption and processing of the data should return to the ecosystem [30]. In [3, 29], the emergence of the concept of DEs is traced and taxonomically situated among related concepts such as Business Ecosystems, Digital Ecosystems, and Platform Ecosystems [16]. Cappiello et al. [6] synthesized the following comprehensive definition of a data ecosystem DE as a 4-tuple DE=⟨Data Sets, Data Operators, Meta-Data, Mappings⟩ where: • Data sets can be structured or unstructured, can have different formats, e.g., CSV, JSON or tabular relations, and can be managed using different management systems. • Data Operators are functions used for accessing or managing data in the data sets. • Meta-Data provides means for describing the DE context domain, can be used to specify the meaning of data and associated data operations. It comprises i) Domain ontology, providing a coherent and unified view of concepts, relationships, and constraints of the domain of knowledge, associating formal semantics with the elements of the domain. If appropriate, several ontologies for different portions of the domain can be devised. ii) Properties that enable the definition of qualitative aspects for the elements of the ecosystem, such as quality and provenance requirements for data sets and operations. iii) Descriptions to associate 111:6 S. Geisler, M.-E. Vidal et al. Fig. 1. A Network of Data Ecosystems Empowered with Strategy and Business Models and Regulations. annotations to the elements of the system for explaining relevant characteristics of data sets and operations. No specific formal language or vocabulary is required in descriptions. • Mappings express correspondences among the different components of the data ecosystem. The mappings are as follows: i) Mappings among ontologies to represent associations among concepts in different ontologies constituting the domain ontology of the ecosystem. ii) Mappings between data sets and ontology to represent relations among the data in the DE data sets and the domain ontology, to allow for their interpretation in terms of the ontology. Data ecosystems can be further empowered with services that exploit the knowledge encoded in the meta-data and operators to satisfy business requirements, such as data transparency and traceability. We nam these knowledge-driven data ecosystems. Services include query processing, data transformation, anonymization, data quality assessment, or mapping generation. The following correspond to examples of notable services: • Concept or mapping discovery: identify a new concept or a new mapping using inductive reasoning and techniques from schema matching, taking into account aspects of uncertainty [15]. Based on the result, the domain ontology and the mappings can be augmented. • Data set curation: identify the best way to keep humans in the loop in order to create a curated version of a data set in a DE (see [1] for limitations of humans in matching). Services can also update the properties of a DE to indicate the provenance of the new curated data sets and manage new generated data from data transformation, analysis, and learning. • Procedure synthesis: construct new procedures out of data operators and other building blocks by composing existing services towards new goals. In complex and evolving systems, it is infeasible to program procedures and even queries without automatic support. Also, exploring repositories and libraries of existing procedures should be possible. In our running example, stakeholders like clinics, insurance companies, and researchers can create their own knowledge-driven DE. Each DE comprises data sets and programs for accessing, managing, and analyzing their data. Interoperability issues across data sets of a DE are solved in a unified view represented in the DE ontology. Mappings between the data sets and the DE ontology describe the meaning of the data sets. Moreover, the description of the data operators enhances data transparency and provides the basis for tracking down the computational methods executed against a DE. Knowledge-driven Data Ecosystems Towards Data Transparency 111:7 To enable collaboration across various knowledge-driven DEs and cope with complex scenarios, a set of DEs can be connected in a network. For this, we envision an ecosystem-wide meta-data layer where the entire ecosystem is described. Figure 1 depicts a network where nodes and edges correspond to knowledge-driven DEs and mappings among them, respectively. In this configuration, the meta-data layer describes each of the nodes (i.e., DE) in terms of descriptions, properties, and domain ontologies. The following mappings can be defined among a network DEs: • Mappings between domain ontologies: state correspondences among the domain ontologies of two nodes or between one node and the global meta-data layer. • Mappings between properties: describe relationships among properties in two nodes. For example, the provenance of two curated versions of a data set could be the same. • Data set Mappings: represent correspondences among data sets of different nodes. Finally, knowledge-driven DEs can be enhanced with additional meta-data layers to enable the description of business strategies and the access regulations. Figure 1 depicts the main components of a network of knowledge-driven DEs empowered with these layers. As can be observed, this enriched version of a network of DEs comprises: (1) Ontological formalisms or causality models that enable the description of the relationships between the data sets received as input and produced as output for the services or operators of a DE. (2) Meta-data describing business strategies that enable the definition of the stakeholders of the network and their roles. (3) Objectives to be met and the dependencies among the tasks that need to be performed to achieve these objectives. (4) Agreements for data exchange and criteria for trustworthiness. (5) Regulations and licenses for data access and for data privacy preservation. (6) Services composing services of the nodes of the network. (7) Services able to monitor and explore decisions taken by services and operators. A network of knowledge-driven DEs will facilitate controlled data exchange across stakeholders in the running example. This network can be hosted and maintained by the consortium of stakeholders. The meta-data layer specifies alignments among the data sets in each DE. It enables the definition of business models and access regulations to be satisfied when the data from one DE (e.g., the clinics DE) is transferred to other DEs (e.g., the DE of the insurance companies or the researchers). Moreover, the formal descriptions of the data sets and operators enhance transparency not only at an individual DE level (e.g., clinics or insurance companies), but also in the network. Lastly, services to monitor how exchanged data is processed in the various multi-site clinical studies empower the network to verify if legal and ethical regulations are fulfilled. 4 ENABLING DATA TRANSPARENCY IN DATA ECOSYSTEMS In this section, we discuss what are the challenges implementing the three groups of requirements introduced in Section 2.2 in the network of knowledge-driven DEs. In Section 4.2 we evaluate and rate how hard it is to meet the requirements for the different types of DEs we defined in this section. 4.1 Challenges of Enabling Data Transparency in Data Ecosystems Data Management Requirements. Sharing heterogeneous data requires guarantees in terms of data quality and transparency. As regards data quality, all the collaborative entities should assess data quality using a common set of DQ services. In fact, in order to get the maximum benefits from the participation to a DE, actors should be able to identify, evaluate, and get the most suitable data for the intended usage. Starting from facilitating the access to available DE data, some existing solutions propose that query processing over heterogeneous data sets rely on a unified interface for overcoming interoperability issues, usually based on meta-models [20]. A few DEs have been proposed, mainly focusing on data ingestion and meta-data extraction and management. Exemplary approaches include Constance [17] 111:8 S. Geisler, M.-E. Vidal et al. and Ontario [13]. Data integration enables the transformation of heterogeneous data sources or views under a unified access schema [25]. Data integration systems comprise data collection and curation steps and resort to record linkage, schema matching, mapping, and data fusion to integrate data from a collection of data sets [9]. At the heart of the data integration realm lies the matching task [5], in charge of aligning attributes of data sources both at a schema and data level, in order to enable formal mappings. Numerous algorithmic attempts (matchers) were suggested over the years for efficient and effective integration (e.g., [7, 11, 22, 24]). Both practitioners and researchers also discussed data spaces as an appropriate data integration concept for DEs. Data spaces do not require a common schema and achieve data integration on a semantic level. Moreover, only a small percentage of data integration systems provide causal explanations to support traceability [34], as well as query processing methods to navigate these explanations [32] efficiently. Existing rule-based approaches that allow for a declarative specification of data transformation, integrity, and integration represent building-blocks for tracking the validity of the domain constraints in all the data-driven pipeline steps in a knowledge-driven DE. These approaches include rule-based entity linking (e.g., [33]), mapping-based tools to perform the process of data integration (e.g., [21]), and declarative languages like SHACL [8], to describe integrity constraints. Knowledge-driven DEs represent a new paradigm for data integration able to trace and annotate provenance and causal relations existing during data ingestion, curation, and integration [18]. DQ challenges are also related to the data variety, and to the fact that the description and measurement of data quality is highly subjective [31], especially if data are used for completely different purposes from the ones they were originally collected for (re-purposing). Algorithms able to describe and assess the quality of very heterogeneous sources and very different stakeholder views are required, together with an agreement about DQ assessment standards in the DE. These standards must include general DQ dimensions but also be derived from the domain at hand to be accepted by the parties sharing the data. The DQ assessment phase requires to provide meta-data and rules to support the selection and reuse of data. DQ should be also assessed on derived data. Challenges here are related to the evaluation of the quality of the outputs of any transformation (e.g., aggregation, formulas, integration) and to the fact that in some cases quality evaluation could be performed in a semi-automatic way. Different stakeholders may have different views on them (e.g., different levels of granularity) and, as a consequence, their integration needs may vary. The input offered by stakeholders regarding the quality of the integration is therefore needed to tailor it to their needs. Data transparency is enabled by a combination of accurate meta-data, including provenance. The facilities to add these to a single, isolated DE are described in Section 3. Networks of DEs introduce further complications: i) meta-data descriptions provided for similar entities by each DE may differ, and may require semantic re-conciliation; ii) levels of transparency provided through provenance may differ across the nodes. A general challenge is therefore to achieve full transparency, or at least formally characterize the boundaries of what is visible, in the presence of łblack spotsž in the global information flow across all nodes in the network. Specific transparency challenges are also related to the difficulties in tracking all the operations performed on data. These new tasks call also for novel reasoning services, based on a sort of reverse engineering process, which rewrites data source queries in terms of global schema (or, ontology) expressions. Future technical support for transparency, traceability, and usage policy enforcement could be developed from distributed ledger technology (e.g., blockchains) and secure multi-party computation (MPC) services. Organization-centric Requirements. The enabling of data governance by several collaborating entities requires the management of different governance models, different levels of data governance maturity, and a clear attribution of responsibilities. The traceability of data sovereignty requires process or dependency models which define how data sovereignty has been executed, e.g., by usage Knowledge-driven Data Ecosystems Towards Data Transparency 111:9 control. The challenge is here to define the content and granularity of information a data owner needs to know based on her role according to regulations valid for the domain. A further challenge is to consider data quality assessment and curation, that can add bias to the data and may also lead to unwanted usage of the data. We need to find business models and regulations which respect the interests of involved stakeholders, adequately react to unwanted data usage, and increase the trust between them. A big challenge is how a data ecosystem can ensure that stakeholders comply to data and data processing standards. Usually, all stakeholders implement their own processes in multiple ways. Enforcing certain standards in a data ecosystem or network of data ecosystems to ensure data quality and data transparency across heterogeneous processes and their interactions is challenging. It requires to negotiate common standards and developing a model feasible for SMEs and big stakeholders to implement. Stakeholders could get a certification which testifies that they adhere to the standards. This makes the quality grade of data sources more transparent to data consumers. As described by Curry and Sheth [10] data ecosystems may vary according to the degree of interaction between stakeholders, coordination of data exchange, and control over data sources. This variety heavily influences also data quality management as it may substantially change the way to assess, monitor, and improve data quality depending on the model. Legal & Ethical Requirements. The satisfaction of laws, fundamental rights, or ethical guidelines demands traceability and certification of data-driven pipelines in a DE. The big challenge is to devise services capable of certifying robustness in data ecosystems according to the national and international legal norms for data protection and fundamental rights, while safeguarding data sovereignty. Furthermore, formalisms, models, and computational algorithms able to interoperate across various stakeholders represents grand challenges to ensure data transparency. 4.2 How can Data Ecosystems Fulfil Data Transparency Requirements? Concluding our analysis, we analyze the three DE architectures presented in Section 3 with respect to their potential in satisfying the requirements and overcoming the challenges towards data transparency. Each requirement is graded following a three-stars scheme: i) ★ means that the requirement is unsatisfied; ii) ★ ★ indicates that the data ecosystem has the potential to satisfy the requirement, but it is challenging; iii) ★ ★ ★ states that the data ecosystem has the potential to fully satisfy the requirement. Each requirement is also evaluated in terms of three levels of satisfaction: I) Complete: The requirement is fully achieved. II) Traceable: The results of the requirement implementation can be traced down. III) Verifiable: The inspection, demonstration, test, and analysis of the requirement implementation can be verified. Table 1 summarizes the analysis of the types of DEs in Section 3. We can observe that the baseline architecture of DEs has the potential to fulfill several requirements (i.e., DMR1-DMR6, OCR4, L&ECR2, and L&ECR5). However, since these DEs are only equipped with data sets, operators, and meta-data, it is challenging for them to keep the stakeholders in the loop during the data quality rating or for assessing the impact of adding new components. For the same reason, these DEs cannot trace or validate the satisfaction of none of the requirements. In contrast, the other two types of DEs are able to fully satisfy the data management requirements (i.e., DMR1-DMR6). Nevertheless, requirement traceability and verifiability still remain a challenge because of the multiple problems of interoperability, data access, and legal regulations imposed by the stakeholders of a knowledgedriven DE or each individual node in a network of DEs. Moreover, a single knowledge-driven DE cannot interact with other DE or circulate their business, regulations, or strategies. As a result, most organization-centric and legal and ethical requirements cannot be satisfied, or if so, it is very challenging. Lastly, networks of knowledge-driven DEs are equipped with meta-data, services, and strategic and business models that facilitate the description of each node and the documentation of 111:10 S. Geisler, M.-E. Vidal et al. Table 1. Three-stars Model for Requirement Satisfaction. DMR1 DMR2 DMR3 DMR4 DMR5 DMR6 Data Ecosystem Complete Traceable Verifiable ★★★ ★ ★ ★★★ ★ ★ ★★★ ★ ★ ★★★ ★ ★ ★★ ★ ★ ★★ ★ ★ OCR1 OCR2 OCR3 OCR4 OCR5 Complete ★ ★ ★ ★★★ ★ Data Ecosystem Traceable Verifiable ★ ★ ★ ★ ★ ★ ★ ★ ★ ★ L&ER1 L&ER2 L&ER3 L&ER4 L&ER5 L&ER6 L&ER7 Complete ★ ★★ ★ ★ ★★ ★ ★ Data Ecosystem Traceable Verifiable ★ ★ ★ ★ ★ ★ ★ ★ ★ ★ ★ ★ ★ ★ Data Management Requirements Knowledge-Driven Data Ecosystem Network of Knowledge-Driven Data Ecosystems Complete Traceable Verifiable Complete Traceable Verifiable ★★★ ★★ ★★ ★★★ ★★ ★★ ★★★ ★★ ★★ ★★★ ★★ ★★ ★★★ ★★ ★★ ★★★ ★★ ★★ ★★★ ★★ ★★ ★★★ ★★ ★★ ★★★ ★★ ★★ ★★★ ★★ ★★ ★★★ ★★ ★★ ★★★ ★★ ★★ Organizational Challenges Knowledge-Driven Data Ecosystem Network of Knowledge-Driven Data Ecosystems Complete Traceable Verifiable Complete Traceable Verifiable ★ ★ ★ ★★★ ★★ ★★ ★ ★ ★ ★★★ ★★ ★★ ★ ★ ★ ★★★ ★★ ★★ ★★★ ★★ ★★ ★★★ ★★ ★★ ★ ★ ★ ★★★ ★★ ★★ Legal & Ethical Challenges Knowledge-Driven Data Ecosystem Network of Knowledge-Driven Data Ecosystems Complete Traceable Verifiable Complete Traceable Verifiable ★ ★ ★ ★★★ ★★ ★★ ★★ ★★ ★★ ★★ ★★ ★★ ★ ★ ★ ★★ ★★ ★★ ★ ★ ★ ★★ ★★ ★★ ★★ ★★ ★★ ★★ ★★ ★★ ★ ★ ★ ★★ ★★ ★★ ★ ★ ★ ★★ ★★ ★★ the negotiations required to exchange across the network. Thus, despite traceability and verifiability are challenging, these DEs are the only ones furnished with components to enable data transparency. We hope that this analysis contributes to the understanding of data transparency challenges. We also aim to encourage the research community to develop trustable networks of knowledge-driven DEs, enabling, thus, DQ management, data governance, and sovereignty, as well as mechanisms to trace and verify the requirement fulfillment. 5 CONCLUSION In this work, we have tackled the challenges that DEs face on their way to become łsmarter," equipped with a knowledge layer. In particular, we focused on data quality and data transparency challenges. Using the motivating example of multi-site clinical studies, we have outlined six data management requirements, five organizational-centric requirements, and seven legal and ethical requirements. We then presented a specific architecture from which data quality challenges were derived and discussed. Table 1 summarizes the discussion by presenting for each of three types of DEs to what extent each of the requirements can be completed, traced, and verified. With the increasing need for integrated data sets and infrastructures to support DEs, we expect their impact on organizations to increase. As data quality in general and data transparency in particular, become a significant issue in data management, we hope this work offers a guideline for researchers and practitioners when investigating developments of knowledge-driven DEs. ACKNOWLEDGEMENTS The authors are grateful to the Dagstuhl team for hosting us in September 2019 (Dagstuhl Seminar 19391). Initial ideas that serve as a basis for this paper were originated and discussed there. Gal also acknowledges the support of the Benjamin and Florence Free Chair. Lenzerini acknowledges the support of MUR-PRIN project łHOPEž, grant n. 2017MMJJRE, and of EU under the H2020-EU.2.1.1 project TAILOR, grant id. 952215. Vidal acknowledges the support of the EU H2020 project iASiS, Knowledge-driven Data Ecosystems Towards Data Transparency 111:11 grant id. 727658 and CLARIFY grant id. 875160. Geisler acknowledges the support of the German Innovation Fund project SALUS, grant id. 01NVF18002. This work has also been supported by the German Federal Ministry of Education and Research (BMBF) in the context of the InDaSpacePlus project (grant id. 01IS17031), Fraunhofer Cluster of Excellence "Cognitive Internet Technologies" (CCIT) and by the Deutsche Forschungsgemeinschaft (DFG) under Germany’s Excellence Strategy EXC-2023 Internet of Production - 390621612. Pernici acknowledges the support of the EU H2020 Crowd4SDG project, grant id 872944. REFERENCES [1] Rakefet Ackerman, Avigdor Gal, and Roee Sagi, Tomerand Shraga. A cognitive model of human bias in matching. In Abhaya C. Nayak and Alok Sharma, editors, PRICAI 2019: Trends in Artificial Intelligence, pages 632ś646, Cham, 2019. Springer International Publishing. ISBN 978-3-030-29908-8. [2] Sebastian R. Bader, Maria Maleshkova, and Steffen Lohmann. Structuring reference architectures for the industrial Internet of Things. Future Internet, 11(7):151, 2019. [3] Martina Barbero, Arne Berre, Davide dalle Carbonare, ..., and Walter Weigel. Towards a European-governed data sharing space. https://www.bdva.eu/sites/default/files/BDVADataSharingSpacesPositionPaperV2_2020_Final.pdf, November 2020. [4] Carlo Batini and Monica Scannapieco. Data and Information Quality - Dimensions, Principles and Techniques. DataCentric Systems and Applications. Springer, Cham, 2016. ISBN 978-3-319-24104-3. doi: 10.1007/978-3-319-24106-7. URL https://doi.org/10.1007/978-3-319-24106-7. [5] Zohra Bellahsene, Angela Bonifati, and Erhard Rahm, editors. Schema Matching and Mapping. Data-Centric Systems and Applications. Springer, Berlin Heidelberg, 2011. ISBN 978-3-642-16517-7. doi: 10.1007/978-3-642-16518-4. URL http://dx.doi.org/10.1007/978-3-642-16518-4. [6] Cinzia Capiello, Avigdor Gal, Matthias Jarke, and Jakob Rehof. Data Ecosystems: Sovereign Data Exchange among Organizations (Dagstuhl Seminar 19391). Dagstuhl Reports, 9(9):66ś134, 2020. ISSN 2192-5283. doi: 10.4230/DagRep.9.9. 66. URL https://drops.dagstuhl.de/opus/volltexte/2020/11845. [7] Chen Chen, Behzad Golshan, Alon Y Halevy, Wang-Chiew Tan, and AnHai Doan. BigGorilla: An Open-Source Ecosystem for Data Preparation and Integration. IEEE Data Eng. Bull., 41(2):10ś22, 2018. [8] Julien Corman, Fernando Florenzano, Juan L Reutter, and Ognjen Savković. Validating SHACL constraints over a SPARQL endpoint. In International Semantic Web Conference, pages 145ś163, Cham, 2019. Springer. [9] Federico Croce, Gianluca Cima, Maurizio Lenzerini, and Tiziana Catarci. Ontology-based explanation of classifiers. In Proceedings of the Workshops of the EDBT/ICDT 2020 Joint Conference, pages 1ś5. CEUR-WS, 2020. [10] E. Curry and A. Sheth. Next-generation smart environments: From system of systems to data ecosystems. IEEE Intelligent Systems, 33(3):69ś76, 2018. doi: 10.1109/MIS.2018.033001418. [11] H. H. Do and E. Rahm. COMA: a system for flexible combination of schema matching approaches. In Proceedings of VLDB, pages 610ś621. VLDB Endowment, 2002. [12] Frederika Welle Donker and Bastiaan van Loenen. How to assess the success of the open data ecosystem? International Journal of Digital Earth, 10(3):284ś306, 2017. doi: 10.1080/17538947.2016.1224938. [13] Kemele M. Endris, Philipp D. Rohde, Maria-Esther Vidal, and Sören Auer. Ontario: Federated query processing against a semantic data lake. In Database and Expert Systems Applications - 30th International Conference, DEXA 2019, Linz, Austria, August 26-29, 2019, Proceedings, Part I, pages 379ś395. Springer, 2019. [14] EU2018. Ethics guidelines for trustworthy AI. https://ec.europa.eu/digital-single-market/en/news/ethics-guidelinestrustworthy-ai, 2018. [15] Avigdor Gal. Uncertain schema matching. Synthesis Lectures on Data Management, 3(1):1ś97, 2011. [16] Tobias Moritz Guggenberger, Frederik Möller, Tim Haarhaus, Inan Gür, and Boris Otto. Ecosystem types in information systems. In Twenty-Eigth European Conference on Information Systems (ECIS2020), pages 1ś21. Association for Information Systems, 2020. [17] Rihan Hai, Sandra Geisler, and Christoph Quix. Constance: An intelligent data lake system. In Proc. of the 2016 International Conference on Management of Data, SIGMOD, San Francisco, USA, pages 2097ś2100. ACM, 2016. doi: 10.1145/2882903.2899389. [18] Pieter Heyvaert, Ben De Meester, Anastasia Dimou, and Ruben Verborgh. Rule-driven inconsistency resolution for knowledge graph generation rules. Semantic Web, 10(6):1071ś1086, 2019. [19] Matthias Jarke. Data sovereignty and the Internet of Production. In Advanced Information Systems Engineering, 32nd International Conference; Grenoble, pages 549ś558, Cham, 2020. Springer. [20] Manfred A. Jeusfeld, Matthias Jarke, and John Mylopouos. Metamodeling for Method Engineering. MIT Press, 2010. 111:12 S. Geisler, M.-E. Vidal et al. [21] Samaneh Jozashoori, David Chaves-Fraga, Enrique Iglesias, Maria-Esther Vidal, and Óscar Corcho. Funmap: Efficient execution of functional mappings for knowledge graph creation. In The International Semantic Web Conference, pages 276ś293, Cham, 2020. [22] B. Kenig and A. Gal. Mfiblocks: An effective blocking algorithm for entity resolution. Information Systems, 38(6): 908ś926, September 2013. [23] F. Kitsios, N. Papachristos, and M. Kamariotou. Business models for open data ecosystem: Challenges and motivations for entrepreneurship and innovation. In 2017 IEEE 19th Conference on Business Informatics (CBI), pages 398ś407. IEEE, 2017. doi: 10.1109/CBI.2017.51. [24] Pradap Konda, Sanjib Das, Paul Suganthan GC, AnHai Doan, Adel Ardalan, Jeffrey R Ballard, Han Li, Fatemah Panahi, Haojun Zhang, Jeff Naughton, et al. Magellan: Toward building entity matching management systems. Proceedings of the VLDB Endowment, 9(12):1197ś1208, 2016. [25] Maurizio Lenzerini. Data integration: A theoretical perspective. In Proceedings of the Twenty-first ACM SIGACTSIGMOD-SIGART Symposium on Principles of Database Systems, pages 233ś246. ACM, 2002. [26] I. Lopez de Vallejo, S. Scerri, and T. Tuikka. Towards a european-governed data sharing space. Technical report, Brussels. BDVA, 2020. [27] Marcelo Iury S. Oliveira and Bernadette Farias Lóscio. What is a data ecosystem? In Proceedings of the 19th Annual International Conference on Digital Government Research: Governance in the Data Age, DG.O 2018, Delft, The Netherlands, May 30 - June 01, 2018, pages 74:1ś74:9. ACM, 2018. [28] Boris Otto and Matthias Jarke. Designing a multi-sided data platform: findings from the international data spaces case. Electronic Markets, 29(4):561ś580, October 2019. URL https://doi.org/10.1007/s12525-019-00362-x. [29] Boris Otto, Dominik Lis, Jan Jürjens, Jan Cirullies, Sebastian Opriel, Falk Howar, Sven Meister, Markus Spiekermann, Heinrich Pettenpohl, and Frederik Möller. Data ecosystems - conceptual foundations, constituents and recommendations for action. Technical report, Fraunhofer ISST, 2019. [30] Rufus Pollock. Building the (open) data ecosystem. https://blog.okfn.org/2011/03/31/building-the-open-dataecosystem/, 2011. [31] Thomas C. Redman. Data Quality: The Field Guide. Digital Press, USA, 2001. ISBN 1555582516. [32] Theodoros Rekatsinas, Sudeepa Roy, Manasi Vartak, Ce Zhang, and Neoklis Polyzotis. Opportunities for data management research in the era of horizontal AI/ML. Proc. VLDB Endow., 12(12):2323ś2324, 2019. [33] Ahmad Sakor, Kuldeep Singh, Anery Patel, and Maria-Esther Vidal. Falcon 2.0: An entity and relation linking tool over wikidata. In CIKM ’20: The 29th ACM International Conference on Information and Knowledge Management, pages 3141ś3148. ACM, 2020. [34] Xiaolan Wang, Laura M. Haas, and Alexandra Meliou. Explaining data integration. IEEE Data Eng. Bull., 41(2):47ś58, 2018. [35] Ruojing Zhang, Marta Indulska, and Shazia W. Sadiq. Discovering data quality problems - the case of repurposed data. Bus. Inf. Syst. Eng., 61(5):575ś593, 2019. doi: 10.1007/s12599-019-00608-0. URL https://doi.org/10.1007/s12599-01900608-0.