BIGOWL4DQ: Ontology-driven approach for Big Data quality meta-modelling, selection and reasoning

Information and Software Technology 167 (2024) 107378 Available online 27 November 2023 0950-5849/© 2023 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/). Contents lists available at ScienceDirect Information and Software Technology journal homepage: www.elsevier.com/locate/infsof BIGOWL4DQ: Ontology-driven approach for Big Data quality meta-modelling, selection and reasoning Cristóbal Barba-Gonzáleza,∗, Ismael Caballerob, Ángel Jesús Varela-Vacac, José A. Cruz-Lemusb, María Teresa Gómez-Lópezc, Ismael Navas-Delgadoa aKHAOS Research group, ITIS Software, Universidad de Málaga, Málaga 29071, Spain bInstitute of Technologies and Information Systems, University of Castilla-La Mancha, Ciudad Real, Spain cIDEA Research group, Dpto. Lenguajes y Sistemas Informáticos, Universidad de Sevilla, Sevilla, Spain ARTICLE INFO Keywords: Data quality evaluation and measurement Data quality information model Big Data Ontology Decision model and notation ABSTRACT Context: Data quality should be at the core of many Artificial Intelligence initiatives from the very first moment in which data is required for a successful analysis. Measurement and evaluation of the level of quality are crucial to determining whether data can be used for the tasks at hand. Conscientious of this importance, industry and academia have proposed several data quality measurements and assessment frameworks over the last two decades. Unfortunately, there is no common and shared vocabulary for data quality terms. Thus, it is difficult and time-consuming to integrate data quality analysis within a (Big) Data workflow for performing Artificial Intelligence tasks. One of the main reasons is that, except for a reduced number of proposals, the presented vocabularies are neither machine-readable nor processable, needing human processing to be incorporated. Objective: This paper proposes a unified data quality measurement and assessment information model. This model can be used in different environments and contexts to describe data quality measurement and evaluation concerns. Method: The model has been developed as an ontology to make it interoperable and machine-readable. For better interoperability and applicability, this ontology, BIGOWL4DQ, has been developed as an extension of a previously developed ontology for describing knowledge management in Big Data analytics. Conclusions: This extended ontology provides a data quality measurement and assessment framework required when designing Artificial Intelligence workflows and integrated reasoning capacities. Thus, BIGOWL4DQ can be used to describe Big Data analysis and assess the data quality before the analysis. Result: Our proposal has been validated with two use cases. First, the semantic proposal has been assessed using an academic use case. And second, a real-world case study within an Artificial Intelligence workflow has been conducted to endorse our work. 1. Introduction Nowadays, the rising volume and the heterogeneity of the types of data made necessary the development of Big Data technologies that facilitate data preparation and Artificial Intelligence (AI) analysis [1]. It is also crucial to define a context where the data quality terms must be specified [2]. To achieve the data quality requirements, data preparation must be performed, but it is not a recipe applied to a dataset; it must be closely related to the specific levels of data quality required for the intended use, depending on the organisation and the moment. Frequently, data scientists consider the actions needed for data quality management only locally for their problem [3]. With ∗Corresponding author. E-mail addresses: [email protected] (C. Barba-González), [email protected] (I. Caballero), [email protected] (Á.J. Varela-Vaca), [email protected] (J.A. Cruz-Lemus), [email protected] (M.T. Gómez-López), [email protected] (I. Navas-Delgado). broader visions of data quality across data for several issues, the efforts achieved to increase data quality can be reduced and shared [4]. For this reason, it is necessary to use a Data Quality Assessment Framework, such as SparkDQ [5], Drunken Data Quality [6], Deequ [7], or Apache Griffin [8], that enables sharing of data quality assessment and improvement results as part of a context-aware Data Preparation Process. However, as we propose in this work, the Data Quality Assessment frameworks do not assess their data quality rules during the design phase to find flaws or inconsistencies before applying them. The wide range of contexts in which the same data can potentially be used makes it necessary to define a standard view of the quality https://doi.org/10.1016/j.infsof.2023.107378 Received 31 March 2023; Received in revised form 17 November 2023; Accepted 26 November 2023 Information and Software Technology 167 (2024) 107378 2 C. Barba-González et al. of data repositories. This is even more necessary due to the wide variety of sources required to feed data repositories, derived from the necessity to integrate various data requirements and formats. Typically, the main activities of the data preparation process are exploration, structure, cleaning, and shaping [9]. One reason why this process is widely recognised as the most time-consuming task for data analysis is because of the lack of shared information throughout the process about the dataset, including the data’s structure and semantics and the data quality levels [10]. In this sense, it is necessary to recall that, without adequate mechanisms, only humans can process raw data, typically coming from various sources with different structures and representation/implementation but with the same meaning. Consequently, actions related to data preparation largely depend on the understanding that humans can extract from the data, including levels of data quality. Thus, it can be said that the mechanism of data preparation depends on the knowledge humans have acquired from observing and using the data. The data preparation process will be benefited from optimising by automating some of these efforts [11]. Business rules are related to data quality in two leastwise ways. First, they let to guide the company’s decisions in its daily operations. Second, they permit the auditing of data produced by existing processes for compliance with external regulations and internal business policies and goals. Thus, business rules offer new perspectives to improve data quality which is essential to have high-quality data to ensure that the results of data analysis are reliable and valuable [12,13]. In this paper, we focus on automating the production and sharing of the data quality assessment results, integrating these results with the metadata of the Big Data workflows, to make the process of applying Artificial Intelligence techniques more efficient. To achieve this goal, we start with BIG data analytics OWL ontology (BIGOWL) [2] and the ontology approach to support knowledge management in Big Data analytics. BIGOWL includes concepts for representing Big Data analytics workflows, including different dimensions: components, workflow definition, and domain knowledge. Details are described in a functional classification that helps them be found when building workflows. However, BIGOWL does not include relevant data management dimensions, such as cybersecurity or data quality. To fill this void, we propose to add the data quality dimension since, as previously said, data quality management must be closed before applying AI solutions (i.e., Machine Learning, Deep Learning, Optimisation). The World Wide Web Consortium (W3C)1proposed the Data Quality Vocabulary (DQ-Vocab),2which provides a metadata model for data quality. However, DQ-Vocab or other existing ontologies [14–18] do not cover the relationship between data quality and data use. Most data quality assessment methodologies [19] directly point to the need to state and validate the corresponding business rules to support the measurement and assessment of data quality. In that regard, it is interesting to note that the Decision Model and Notation (DMN) [20] supports the modelling and evaluation of business rules, described using an expressive language called S-FEEL (Simplified Friendly Enough Expression Language) that combines fundamental expressions [21,22]. Using DMN tables, there are frameworks [23–25] for modelling and evaluation of data quality rules that follow the DMN standard. DMN4DQ a framework that uses the rules to represent the user requirements for the quality of the data in the context of use to generate a recommendation on the usability of the data is presented in [25]. This recommendation is based on evaluating the quality of the data for this context of use. However, in this framework, the rules are not evaluated to identify flaws or inconsistencies before they have been executed. For this reason, we propose the creation of the BIGOWL4DQ ontology, whose main aim is representing and consolidating data quality knowledge for Big Data analytics. 1https://www.w3.org/. 2https://www.w3.org/TR/vocab-dqv/. BIGOWL4DQ extends BIGOWL with data quality dimensions, which provide mechanisms for improving the reasoning capacity and facilitate the incorporation of data quality tasks in Big Data workflows. Thus, our threefold objectives are (1) to represent actionable knowledge to automate the process of better-supporting data preparation; (2) to reduce the complexity of defining the main concepts of Data Quality management within a context-aware Data Preparation Process; and (3) to assess the correctness and completeness of the data quality rules. The challenge of guaranteeing the accuracy of data quality rules, specifically identifying inconsistent or incomplete rules. For example, DMN decision tables are raised by using DMN decision tables as a specification vehicle for crucial business decisions [26]. Reasoning employing BIGOWL4DQ annotations helps identify errors in the rules at the specification stage, which may help avoid costly flaws later on during the design and execution of business processes. We complement the OWL 2 axioms with SWRL (Semantic Web Rule Language) [27] rules for the reasoning process. 1.1. Contributions The main contributions of this study are: •The proposed ontology, BIGOWL4DQ, which has been designed and implemented for a comprehensive solution that integrates and appropriately relates the measurement, assessment, and creation of a suggestion for using data through business rules. •A semantic approach has been modelled and implemented to annotate all the meta-data involved from business rules for defining the main concepts of Data Quality measurement, management, and Data Quality assessment. •BIGOWL4DQ has been designed to include SWRL rules for evaluating the correctness and completeness of business rules for data quality in measurement, assessment, and usability decisions. •The semantic model is evaluated in two contexts, one from an academic use case and the other from a practical use case of smart farms related to soil sensor networks. 1.2. Structure of this work The rest of the paper is structured as follows: Section 2introduces the background needed to understand the proposal and the related work; Section 3details the semantic method required to assess the data quality using the case study from [25] to illustrate the proposal; Section 4presents the application of our proposal to the case study; in Section 5we point out the main threads to validity of this work; Discussion and conclusions are presented in Section 6. 2. Foundations This section introduces the theoretical foundations required to develop the investigation introduced in this manuscript. There are two important groups: one aimed at introducing concepts related to the representation of ontologies and the other dedicated to the field of data quality measurement. A review of the state-of-the-art is also given to highlight the main differences between the related works and the proposed approach. 2.1. Background concepts on ontology and representation of knowledge An ontology offers a formal representation of the real world [28]. It provides properties of each idea, which describe numerous aspects and qualities of concepts (classes of concepts), limitations on properties, and an explicit explanation of concepts in a domain of discourse. A knowledge base comprises an ontology and a collection of unique instances of classes, and it provides services to encourage interoperability across various heterogeneous systems and databases [29]. Information and Software Technology 167 (2024) 107378 3 C. Barba-González et al. OWL is an ontology language based on decidable subsets of firstorder logic: Description Logics (DL). As for DL in OWL, we can differentiate between the T-Box (concepts, relationships and constraints) [30] and the A-Box (relations between individuals and concepts). OWL as a Knowledge Representation Model enables automated inference processes to extract implicit knowledge from a knowledge base. The use of OWL also enables homogenisation in the description process using the explicit knowledge represented, easing the integration of different datasets. The OWL 2 Web Ontology Language, informally OWL 2, is the latest W3C Recommendation for ontologies. There are some OWL 2 editors, but the most well-known is Protégé.3It can be used as a standalone application or a Web-based tool, enabling the team development of OWL ontologies. Semantic Web Rule Language(SWRL)4is a W3C proposal to combine OWL 2 with the Rule Markup Language (RuleML). However, it only applies to the OWL DL subset and the Unary/Binary Datalog RuleML sublanguages. SWRL can be evaluated using existing OWL 2 Reasoners. Thus, this language increases the reasoning capabilities of OWL 2 with production rules. OWL-based ontologies are given procedural knowledge by the Semantic Web Rule Language (SWLR), which makes up for some of the shortcomings of ontology inference, notably in recognising semantic links between instances [31]. SWRL includes a high-level abstract syntax for Horn-like rules.5Model-theoretic semantics provides the formal meaning of OWL 2 ontologies, including rules written in this abstract syntax. An OWL 2 ontology is made up of a list of rules and facts. Each rule has an antecedent and a consequent, where if the antecedent is true, then the consequent must be satisfied. Facts are rules without an antecedent. SWRL provides the typical logic expression ‘‘𝐴𝑛𝑡𝑒𝑐𝑒𝑑𝑒𝑛𝑡 ⇒ 𝐶𝑜𝑛𝑠𝑒𝑞𝑢𝑒𝑛𝑡’’ to represent semantic rules. The antecedent of the pair (rule body) and the consequent (rule head) can be conjunctions of one or more atoms written as ‘‘𝑎𝑡𝑜𝑚1∧𝑎𝑡𝑜𝑚2∧⋯∧𝑎𝑡𝑜𝑚𝑛’’. Each atom has one or more associated parameters, which are denoted by a question mark and a variable (e.g., ?𝑥). The typical uses of SWRL include the transfer of characteristics and inferring the existence of new individuals [32]. Both TBox and ABox are used in the reasoning processes of OWL 2. As in the case of ontology editors, many tools are available to perform querying, inference, and reasoning tasks on ontologies. Reasoners aim at finding implicit knowledge in the ontologies by applying a set of reasoning mechanisms, such as: •Satisfiability: Check whether an ontology class can be instantiated. •Subsumption: Check whether any ontology constraint is implied by the rest of the constraints. •Classification: –Calculate the set of subclasses of each class explicitly, i.e. determine whether there are implicit subclasses. –Calculate whether two concepts are synonymous. –Calculate the most specific class to which an instance belongs. •Consistency: Check whether the instances defined within an ontology satisfy all the restrictions. As stated before, in this work, we extend with data quality concepts the ontology BIGOWL, whose main aim is to provide a broad vocabulary of terms related to Big Data analytics processes, including their components and how they are integrated, from data sources to analytics visualisation. This ontology was created using the OWL 2 ontology language, which uses classes to represent concepts and data attributes 3https://protege.stanford.edu/. 4https://www.w3.org/Submission/SWRL/. 5https://www.w3.org/2005/rules/wg/wiki/Horn_Rules_Semantics.html. Table 1 Data quality dimensions by Wang et al. [39]. Data quality dimensions are the criteria for defining data quality category. Data quality dimensions indicate concepts, principles, and procedures for describing, measuring, analysing, and improving the critical aspects of data. Each row of the table represents the category and its dimensions. Data quality category Data quality dimension Intrinsic Accuracy, objectivity, believability, reputation Accessibility Access, security Contextual Relevancy, value-added, timeliness, completeness, amount of data Representational Interpretability, ease of understanding, concise representation, and consistent representation Table 2 Inherent data quality characteristics from ISO 25012 [40]. A generic data quality model that applies to structured data stored in an information system is defined by the ISO 25012 standard. It establishes five basic quality dimensions common to any standard: accuracy, completeness, consistency, credibility and currentness. Inherent data quality characteristics Definition Accuracy The degree to which data have attributes that correctly represent the true value of the intended attribute of a concept or event in a specific context of use. Completeness The degree to which subject data associated with an entity has values for all expected attributes and related entity instances in a specific context of use. Consistency The degree to which data has attributes that are free from contradiction and are coherent with other data in a specific context of use. Credibility The degree to which data has attributes that are regarded as true and believable by users in a specific context of use. Currentness The degree to which data has attributes that are of the right age in a specific context of use or object properties to describe relations. BIGOWL includes 184 classes, 66 individuals, 488 axioms, 16 object properties, and 20 data properties (individual attributes). In addition, BIGOWL also supports reasoning thanks to the formulation of semantic rules to deduce new information from existing knowledge. These rules are formulated in SWRL and are used to perform semantic reasoning jobs mainly devoted to checking the consistency of workflows. Reasoning tasks are evaluated in this work using Pellet [33], an open-source Java based OWL 2 reasoner. 2.2. Data quality background Data quality is generally understood as fitness for use [34]. There are frameworks and proposals [5–8,23–25] for assessing data quality. They usually focus on satisfying data quality as the fundamental guarantee for data-based research, decision-making, and service [35– 37]. However, to judge data quality in a given context [38], one or more criteria are necessary to evaluate or assess the quality. These criteria are commonly known as data quality dimensions in the scientific literature [39] or data quality characteristics in the context of ISO standards [40]. The set of various data quality characteristics that are useable and eligible (e.g., to describe the most representative user data quality requirements) is called data quality model. Over the years, several authors and practitioners have provided their own data quality model adapted to their specific context of use. However, two of the most widely used and referenced are the one proposed by [39] (see Table 1) or the one proposed by [40] (see Table 2). It is necessary to deploy a specific measurement method to determine whether a dataset has an adequate level of quality for an exact data quality dimension or characteristic representing a particular data quality requirement. The measurement method typically depends on the technology in which the data set is implemented [25,41]. Furthermore, to calculate the measurement, a specific set of business rules must be Information and Software Technology 167 (2024) 107378 4 C. Barba-González et al. Fig. 1. Data Quality Measurement Information Model (DQMIM) [44] the referred terms by following ISO/IEC 15939 as a basis. These terms refer to the data quality employing concepts as an entity, attribute, and measurable concept. conveniently selected and grouped for all data quality chosen characteristics [42]. The measurement result, typically calculated as the ratio of data records that do not violate the stated business rules and the total number of records, must then be compared with the specific threshold values representing the organisation’s risk appetite for the task at hand. If the comparison is unfavourable, data quality analysts must decide on discarding or cleaning the data. Besides, depending on how bad the achieved measurement is obtained and the viability of the operation or the cost of cleaning the data is disproportionate, it may be easier to recapture or refuse to use the data. Thus, depending on the violations of business rules, specific actions must be designed and executed for data cleaning [43]. One of the most challenging works in the literature has been to represent the related knowledge that an organisation can manage about the levels of data quality of their data repositories. Apart from the specific sets of concepts provided in ISO/IEC 25012, ISO/IEC 25024, at this point, we would like to highlight two seminal works: DQMIM [44] and Vocab-DQ.6DQMIM, which stands for Data Quality Measurement Information Model (DQMIM), is based on the ISO/IEC 15939 standard [45] and provides a set of terms related to the measurement of software quality. These terms were adapted for the data quality, such as entity, attribute, and measurable concept (see Fig. 1). The main limitation of this model concerning the proposal of this work lies in the fact that, although it is sufficiently complete for a high level of abstraction, it does not allow for the incorporation of concrete details of the implementation of the measures. In 2016, the W3C Working Group published Vocab-DQ, a Data Quality Vocabulary, to gather and align some previous work in the area with the ISO/IEC 25012 standard [40]. Their main aim was to provide a framework in which the quality of the datasets could be adequately described (see Fig. 2). Vocab-DQ is one of the most important contributions to representing data quality concepts using semantic web technologies. However, this vocabulary does not have the semantic capacity to express the relationship between the measurement and evaluation concepts described in DMN4DQ framework. 2.3. Rule modelling language for data quality description The definition of data quality requirements through a set of rules facilitates modelling and a later evaluation with the tools that support 6https://www.w3.org/TR/vocab-dqv/. the existing rule-modelling languages. Various standards support the modelling, description, and evaluation of business rules, such as the following: •SBVR7defines the vocabularies and rules required to communicate organisations and software tools due to the definition of common elements. SBVR has a sound theoretical foundation of formal logic: it is based on first-order predicate logic with extensions into modal logic, i.e., some deontic forms for expressing obligation and prohibition and alethic forms for expressing necessities and possibilities [46]. •Semantic Web Rule Language (SWRL)8combines the OWL DL and OWL Lite sublanguages. SWRL includes a high-level abstract syntax for Horn-like rules. Additionally, model-theoretic semantics is given to provide the formal meaning for OWL ontologies, including rules in this abstract syntax. An OWL ontology is made up of a list of rules and facts. Each rule has the antecedent, and consequent, where if the antecedent is true, the consequent must be satisfied. Facts are rules without an antecedent. •Production Rule Representation (PRR)9is an OMG standard to provide a vendor-neutral rule-model representation in UML for production rules. The standard provides a set of metamodels and OCL restrictions to define production rules enabling the afterwards transformation of those models to any rule engine. •Decision Model and Notation (DMN)10 is an OMG standard to obtain and represent decision models through a declarative description of the form ‘if then’ [47]. It is a standard notation for capturing decision logic that can be used in general business applications. DMN facilitates the modelling of repeatable decisions according to the necessities and is supported by a set of engines, such as Camunda - DMN Engine11 or Drools - DMN.12 DMN has also been used to represent data quality rules and validated in real datasets [20,25]. •RuleML (Rule Markup Language)13 represents a family of languages for the specification of rules in the web context. RuleML can be used as a bridge between other OMG languages, such as SWRL, SBVR or PRR. There exist translators between RuleML and DMN languages, such as [48]. 2.4. Related work The use of semantics in the area of data quality management can be summarised in the following major uses [49]: •Collaborative representation and use of quality-relevant knowledge. The semantics facilitates the definition of the data requirements in a structured and shareable way. Besides, semantics enables a machine-processable data format that is also readable by humans. This representation focuses on defining vocabularies [50], data dependencies [51], or rules [52]. •Automated identification of conflicting data requirements into harmonised data. For example, [53] focuses on using semantics to manage incomplete terminologies, irrelevant terms, outliers, missing values, data categorisation, and duplicated terms to make data standardisation in the medical data context. 7https://www.omg.org/spec/SBVR/1.5/About-SBVR/. 8https://www.w3.org/Submission/SWRL/. 9https://www.omg.org/spec/PRR/. 10 https://www.omg.org/dmn/. 11 https://camunda.com/products/dmn-engine/. 12 https://www.drools.org/learn/dmn.html. 13 http://ruleml.org/. Information and Software Technology 167 (2024) 107378 5 C. Barba-González et al. Fig. 2. Data Quality Vocabulary (DQV) of W3C. DQV provides a framework for describing the quality of a dataset. To express these qualities about a dataset, employ five different types of quality information represented by the following classes: dqv:QualityAnnotation for describing feedback and quality certificates, dcterms:Standard whose main is to represent a standard the dataset or its distribution. dqv:QualityPolicy describes a policy governed by data quality concerns. dqv:QualityMeasurement whose goal is to provide a metric value about the dataset or distribution that delivers quantitative or qualitative information. And finally, prov:Entity describes an entity involved in the provenance of the dataset or distribution. Furthermore, DQV disposes of concepts and properties for data dimension, metrics, category, etc. •Semantic definition of data. Ontology elements support the creation of textual descriptions of the semantics of their classes and properties. Therefore, data providers can identify concepts from other ontologies that can be reused for their own data. For example, in [54], the authors explain how to extract semantic features of images to define their data. •Use of semantics data as reference data. Semantics provides an enormous variety of data from several domains, such as biology, media, and life sciences, which can be used as trusted reference data in data quality monitoring. For example, [50] describes the requirements for legal value and functional dependency in any domain. This work proposes using RDF to format the data. Thus, semantics can be employed to compare data with the trusted reference data. •Content integration with ontologies. Defining hierarchies and class relationships in ontologies ease content analysis at several levels, satisfying their definitions. Besides, the modelisation of equivalence relationships explicitly helps to analyse the data without knowing all synonym relationships. For instance, [55] defined conceptual models in a flexible and extendable way, which allows modelling complex analytical tasks for heterogeneous data sources. Regarding data quality ontologies, they have been the subject of extensive research in recent years, as data quality has become increasingly important in various domains, such as healthcare [16,56], ground [18], finance [17], and government [57]. Moreover, [58] presents a methodology and an ontology to assess the data quality, but only when the data is in RDF format. Furthermore, other approaches help to represent data quality assessment, such as data Quality Management (DQM) [52,59], Data Cleaning Ontology (DCO) [60], Data Quality Ontology (daQ) [61], Data Quality Vocabulary(DQV) [62] and Fuzzy Quality Data Vocabulary (FQV) [63]. The ontologies above do not help assess the data quality; instead, they publish quality reports in a machine-readable manner. From a different point of view, Reasoning Violation Ontology (RVO) [61] is a dedicated reasoning error ontology that allows processing data issues. [64] evaluates data quality employing semantic rules which can be user-defined in an ontology for data stream applications. In addition, [15] defines concepts and data quality measures in healthcare data. [65] presents an ontology-based data quality framework for relational data stream management systems that include data quality measurement; this work briefly describes data quality concepts through an ontology. Linked Open Data (LOD) [66] and ontologies are used to describe data quality characteristics due to LOD enables defining the presence of interlinks between datasets and using ontologies as data schemes [67]. LOD allows facing data quality essential issues such as missing data, missing entity relationships, and erroneous data values [68]. Furthermore, LOD can transform data from one format to linked data. However, this transformation can degrade data quality due to various problems, such as errors introduced at the source, parsing values, or interpreting [69]. Data integration in LOD from multiple sources does not continually improve data quality due to contradictory information from the different sources [70]. Regardless of the total number of integrated data sources, quality issues persist at the schema and instance levels [71]. However, many ontologies are inaccessible or are only sparsely described in published scientific works. In addition, the available ontologies do not fulfil our needs completely of creating an ontology for incorporating data quality dimensions, which provides mechanisms for conducting reasoning over the data quality annotations and facilitates the incorporation of the data quality tasks in Big Data workflows. Information and Software Technology 167 (2024) 107378 6 C. Barba-González et al. Fig. 3. Example of the Business Rule in a DMN table used in Data Quality Assessment. Each row is a rule describing the evaluation of the data quality by a set of Business Rule Data Quality Measurement (BR.DQM) combining the results of the measurement of several data quality dimensions. Fig. 4. Overview of the relations of the BIGOWL4DQ ontology. BIGOWL4DQ reuses concepts related to Data Quality and Big Data created previously. Thus, it is aligned with Data Quality Vocabulary, BIGOWL, and with the standards ISO 25012 and 25024. 3. Semantic model of BIGOWL4DQ One of the main goals of this study is to represent data quality measurement, management concepts, and relationships in Big Data analysis using a formalised language supported by an engine to evaluate data quality rules. For this reason, we opted to extend BIGOWL ontology with the expressivity of DMN4DQ to describe a set of data quality rules in a defined domain and context. To the best of our knowledge, DMN4DQ is the only existing methodology supported by an engine that can assess the data quality of a dataset, based on a set of data quality rules that are machine-readable and processable. This is the key reason why we have based our BIGOWL extension on DMN4DQ. The proposal of DMN4DQ [25] describes a hierarchical structure which is defined (from bottom to top) as follows: (i) the Business Rule for Data Value (BR.DV) evaluates every data record provided as input of the dataset; (ii) for each data quality characteristic, a Business Rules for Data Quality Measurement (BR.DQM) combines the retrieved outputs of the required BR.DVs as input to generate a data quality measure; (iii) Business Rules for Data Quality Assessment (BR.DQA) uses the outputs of different DMN tables related to the measurement of a dimension (BR.DQM) as input to generate a data quality assessment as shown in the example of Fig. 3; and, (iv) BR.DUD takes the outputs of BR.DQA as input determines the level of usability of each record. For sharing the data quality rule vocabulary and the data of the Big Data workflow, we propose BIGOWL4DQ. This new ontology is aligned with the W3C Data Quality Vocabulary, which provides an extended group of concepts regarding the quality of the dataset. It also reuses the standards ISO 25012 and 25024 as shown in Fig. 4. Thus, BIGOWL4DQ includes the TBox to represent data quality measurement, management concepts, and relationships. Therefore, these elements can be used later to introduce specific use cases through ABox. Furthermore, our ontology provides a series of Semantic Web Rule Language rules to validate the correctness and completeness of the business rules. 3.1. BIGOWL4DQ: An ontology for data quality management The main goal of this work is to capture all the semantics needed to define data quality measurement and assessment. For this reason, we opt to design a new OWL 2 ontology extending BIGOWL to describe business rules, data quality characteristics, and quality measures, among others, in the Big Data context. To this end, the standard Ontology 101 development process [28] has been followed, which comprises seven steps: 1. Determine the domain and scope of the ontology. The main scope of BIGOWL4DQ is the evaluation and assessment of data quality in big data environments. This scope involves business rules for data values, data quality measurement, assessment, or usability decisions orientated to Big Data. 2. Consider reusing existing ontologies. The proposed ontology extends BIGOWL, which has been successfully assessed to define the lifecycle of a workflow, from data reading to the results view. Furthermore, BIGOWL4DQ reuses DQ-Vocab concepts, which provide a metadata model for data quality, e.g. Dimension, Dataset, etc. 3. Enumerate important terms in the ontology. Outstanding terms have been selected from the literature on data quality, specifically DQMIM and DQ-Vocab. In addition, terms from the ontologies aligned [62] are incorporated as well as Data Quality Vocabulary. Examples of such terms are QualityMeasurement or DataSet. 4. Define the classes and the class hierarchy. We have followed a top-down approach in developing the class hierarchy. This fact makes it easier to align with BIGOWL and Data Quality Vocabulary, create annotation mappings, and employ a semantic reasoner, among other things. Fig. 5 shows the core classes of the ontology and the hierarchy of the ontology. For example, class Business Rules has several subclasses, including Business Rules Data Usability Decision,Business Rules Data Quality Assessment, Business Rules Data Quality Measurement, and Business Rules Data Value. BIGOWL4DQ has been developed using Protégé and OWL 2. 5. Define the properties of classes and slots. 37 object properties and 24 data properties have been included to relate classes and define attributes. An illustrative set of properties is shown in Table 3, where the class Clause is related to the class Condition by means of the property of the object hasCondition. Data properties of class Clause are annotation, or hasAssociatedHitPolicyCriteria. 6. Define the facets of the slots. This step aims to include cardinality constraints and value restrictions for the ontology’s properties. For example, the range of the property creationDate is restricted to the date to indicate when the class BussinesRule is in its domain. 7. Create instances. The instances or individuals in BIGOWL4DQ are specific to the data quality domain. For example, Completeness is an instance of the class DataQualityCharacteristic. The class BusinessRule has a property hasHitPolicy (with range Policy) to indicate general assertions or guidelines on how a business rule is intended to operate. For example, to validate our proposal, an educational use case from [25] was used to create its rules as instances of BIGOWL4DQ and then assessed them. For example, Fig. 6 shows a partial view of the RDF Graph for the DMN tables for BR.DV.04. The complete set of ontology instances of this case is available at ‘‘Case-study-data-quality-DMN4DQ.owl’’ file at the GitHub repository.14 Fig. 5 shows the hierarchy to describe the primary data quality ideas. There are defined classes to represent, throughout business rules, 14 https://github.com/ProyectoAether/BIGOWL4DQ. Information and Software Technology 167 (2024) 107378 7 C. Barba-González et al. Fig. 5. Overview of the BIGOWL for Data Quality ontology. Subclasses are shown by continuous arrows, whereas dotted arrows denote characteristics. BIGOWL4DQ defines data quality concepts. There are four types of business rules: Data Value, Data Quality Measurement Dimension, Data Quality Dimension, and Data Usability Dimension. Furthermore, it describes Quality Measure concepts, clauses, methods, decision criteria, conditions, etc. Table 3 Data and object properties to annotate data quality concepts. It can be specified features like business rule domain or clause, hit policy, etc. Object properties Description logic Textual description domain ∃domain Thing ⊑BusinessOutput⊤ ⊑ ∀domain (PrimitiveType ⊔BusinessOutput) Specifies the domain of the business output, ensuring it belongs to the PrimitiveType or BusinessOutput category. generates ∃generates Thing ⊑QualityMeasurementWay⊤ ⊑ ∀generates QualityMeasure Represents the method or way by which quality measurements are generated, falling under the QualityMeasurementWay or QualityMeasure categories. hasAssociated ∃hasAssociated Thing ⊑Dataset⊤ ⊑ ∀hasAssociated DataUsabilityRecomendation Links the property to its associated dataset, falling under the Dataset or DataUsabilityRecommendation category. hasClause ∃hasClause Thing ⊑BusinessRule⊤ ⊑ ∀hasClause Clause Establishes a connection between a business rule and its associated clause. It ensures that clauses belong to the BusinessRule or Clause category. hasCondition ∃hasCondition Thing ⊑Clause⊤ ⊑ ∀hasCondition Condition Indicates the conditions associated with a clause, ensuring they belong to the Clause or Condition category. hasCriteriaDecision ∃hasCriteriaDecision Thing ⊑BusinessRuleDataUsabilityDecision⊤ ⊑ ∀hasCriteriaDecision DecisionCriteria Connects a business rule to its data usability decision criteria, ensuring they fall under the BusinessRuleDataUsabilityDecision or DecisionCriteria category. involves ∃involves Thing ⊑QualityMeasurementWay⊤ ⊑ ∀involves DecisionCriteria Links a property to the decision criteria involved in quality measurement, ensuring they belong to the QualityMeasurementWay or DecisionCriteria category. Data properties Description logic Textual description annotation ∃annotation Datatype ⊑Clause⊤ ⊑ ∀annotation Describes additional information or metadata associated with a clause, falling under the Datatype or Clause category. businessRulesSource ∃businessRulesSource ⊑BusinessRule⊤ ⊑ ∀businessRulesSource Specifies the source of business rules, ensuring they belong to the BusinessRule category. creationDate ∃creationDate ⊑BusinessRule ⊤ ⊑ ∀creationDate Datatype Represents the date when a business rule was created, falling under the BusinessRule or Datatype category. hasAssociatedHitPolicyCriteria ∃hasAssociatedHitPolicyCriteria ⊑Clause⊤ ⊑ ∀hasAssociatedHit- PolicyCriteria Datatype Links a property to its associated hit policy criteria, falling under the Clause or Datatype category. name ∃name ⊑Field Represents the name associated with a field. isCorrect ∃isCorrect⊑BusinessRule⊤ ⊑ ∀isCorrect Datatype Indicates whether a business rule is correct, falling under the BusinessRule or Datatype category. isCompleteField ∃isCompleteField⊑Field⊤ ⊑ ∀isCompleteField Datatype Specifies whether a field is complete, falling under the Field or Datatype category. isCompleteCondition ∃isCompleteCondition⊑Condition⊤ ⊑ ∀isCompleteCondition Datatype Indicates whether a condition is complete, falling under the Condition or Datatype category. isCompleteClause ∃isCompleteClause⊑Clause⊤ ⊑ ∀isCompleteClause Datatype Specifies whether a clause is complete, falling under the Clause or Datatype category. isComplete ∃isComplete⊑BusinessRule⊤ ⊑ ∀isComplete Datatype Indicates whether a business rule is complete, falling under the BusinessRule or Datatype category. the user requirements for the data in a context of use to generate a recommendation on the usability of the data. There are four types of business rules: Data Value, Data Quality Measurement Dimension, Data Quality Dimension, and Data Usability Dimension. To measure the quality of the different business rules, the class Data Quality Characteristic, measured by the class Quality Measure, uses different ways annotated with the class Quality Measurement Way and methods determined by the class Calculation Method. Furthermore, more classes Information and Software Technology 167 (2024) 107378 8 C. Barba-González et al. Fig. 6. General overview of an example of BIGOWL4DQ instances for one DMN table. In red, the ellipses depict the classes described in our ontology. In addition, in black ellipses, the samples created to illustrate the example of the DMN table are shown. The arrows represent the properties of the data and objects. are indicated to describe other concepts like clauses, decision criteria, conditions, etc. The complete ontology is developed in ‘‘bigowl4dq.owl’’ file and available in the GitHub repository14. 3.2. Reasoning rule framework The high number of business rules for data quality can imply a complex validation process. Some of the most common validations are based on analysing correctness and completeness. In the context of DMN tables, Dumas et al. [72] proposed a formal validation that analyses possible missing and overlapping rules, duplicate rules, conflicting rules, shadowed rules, types of expression, correct use of enumerations, and correctly connected requirement graphs. The proposal is supported by two tools to validate the DMN tables: dmn-js15 and dmn-check.16 In contrast with those tools, our proposal validates DMN tables and any business rules. This section describes a set of SWRL rules to control correctness and completeness using a semantic-based approach. The SWRL rules can consider the data quality rule aspects and the semantic context of the problem under study. BIGOWL4DQ defines a set of SWRL rules on top of the OWL 2 ontology to derive new information from existing knowledge. These rules are used by semantic reasoning tasks primarily concerned with verifying the validity of the rules, such as identifying rules that are (in)compatible with the input domain, clauses, conditions, etc. Another possible use is to check the completeness of the rules, i.e., every potential input configuration trigger has at least one associated rule. Thus, the primary purpose of these rules is to create well-formed business rules. These SWRL are included in the complete ontology (‘‘bigowl4dq.owl’’) available in the GitHub repository14. 15 dmn-js: http://dmn.cs.ut.ee. 16 dmn-check: https://github.com/red6/dmn-check#validations. The following is a list of these rules to check the validity of several concerns of the identified business rules for data quality evaluation and assessment: - SWRL Rule for checking the correctness in Business Rules Data Value. This rule is used to check the compatibility between the domains of the business rules and the field in which the data are applied. In addition, it is assessed that the rules contain clauses and define their conditions and outputs. Correctness of Business Rule Data Value bigowl4dq:BusinessRuleDataValue(?br) ^ bigowl4dq:hasClause(?br, ?cl) ^ bigowl4dq:hasCondition(?cl, ?cond) ^ bigowl4dq:outputClause(?cl, ?outCl) ^ bigowl4dq:inputRef(?br, ?ref) ^ bigowl4dq:hasFieldDomain(?ref, ?indo) ^ bigowl4dq:outputDomain(?br, ?od) ^ bigowl4dq:inputDomain(?br, ?indo) ^ bigowl4dq:outputResult(?br, ?ores) ^ bigowl4dq:domain(?ores, ?od) -> bigowl4dq:isCorrect(?br, true) - SWRL Rule for checking the correctness in Business Rules Data Quality Measurement Dimension. This rule examines the harmony between the input data domain of the Data Quality Measurement Dimension (DQMD) rules and the output data domain of the Data Value rules. Both domains must match because the output of the data quality rules for data values (BR.DV) is the input of the DQMD rules. In addition, it checks that the clauses, conditions, and outcomes are well-defined in the DQMD rules. Information and Software Technology 167 (2024) 107378 9 C. Barba-González et al. Correctness in Business Rules Data Quality Measurement bigowl4dq:BusinessRuleDataValue(?br) ^ bigowl4dq:outputResult(?brDV, ?or) ^ bigowl4dq:outputDomain(?br, ?od) ^ bigowl4dq:inputDomain(?br, ?or) ^ bigowl4dq:hasClause(?br, ?cl) ^ bigowl4dq:hasCondition(?cl, ?cond) ^ bigowl4dq:outputClause(?cl, ?outCl) -> bigowl4dq:isCorrect(?br, true) - SWRL Rule for checking the correctness in Business Rules Data Quality Assessment. This rule checks for compatibility between the input data domain of the Data Quality Assessment (DQA) rules and the DQMD’s output data domain. The result of the DQMD rules is the input of the DQA rules. Therefore, both domains must be compatible. It also ensures that the clauses, conditions, and outcomes of the DQA rules are well-defined. Correctness in Business Rules Data Quality Assessment bigowl4dq:BusinessRuleDataQualityAssessment(?br) ^ bigowl4dq:outputResult(?br, ?or) ^ bigowl4dq:BusinessRuleDataUsabilityDecision(br) ^ bigowl4dq:outputDomain(?br, ?od) ^ bigowl4dq:inputDomain(?br, ?or) ^ bigowl4dq:hasClause(?br, ?cl) ^ bigowl4dq:hasCondition(?cl, ?cond) ^ bigowl4dq:outputClause(?cl, ?outCl) -> bigowl4dq:isCorrect(?br, true) - SWRL Rule for checking the correctness in Business Rules Data Usability Decision. This rule verifies that the input data domain of the Data Usability Decision (DUD) rules and the output data domain of the DQA are compatible. Both domains must be identical because the output of the DQA rules equals the input of the DUD rules. It also adequately defines the DUD rules’ clauses, conditions, and results. Correctness in Business Rules Data Usability Decision bigowl4dq:BusinessRuleDataUsabilityDecision(?br) ^ bigowl4dq:outputResult(?br, ?or) ^ bigowl4dq:outputDomain(?br, ?od) ^ bigowl4dq:inputDomain(?br, ?or) ^ bigowl4dq:hasClause(?br, ?cl) ^ bigowl4dq:hasCondition(?cl, ?cond) ^ bigowl4dq:outputClause(?cl, ?outCl) -> bigowl4dq:isCorrect(?br, true) - Rules for checking the completeness in a condition. This family of rules proves that the range of values defined in a clause condition covers any input value. There are rules for checking the interval of numeric, Boolean, or String values. For instance, the following rule verifies that a maximum value of an interval of a numeric field is covered; it is checked that the maximum value defined in the interval of a condition is equal to the value specified in the other conditions. This means that the first condition defines its interval as [1,3] and the second condition as >3. As a consequence, all values greater than or equal to 1are covered by those conditions. Rules have been described for all different combinations. Case 1: Completeness in a condition bigowl4dq:BusinessRuleDataValue(?br) ^ bigowl4dq:inputRef(?br, ?inRef) ^ bigowl4dq:hasClause(?br, ?cl) ^ bigowl4dq:hasCondition(?cl, ?cond) ^ bigowl4dq:inputDomainCondition(?cond, ?inRef) ^ bigowl4dq:hasValueCondition(?cond, ?valCond) ^ bigowl4dq:inputDomainCondition(?cond2, ?inRef) ^ bigowl4dq:hasValueCondition(?cond2, ?valCond2) ^ bigowl4dq:hasOperator(?valCond, bigowl4dq:BetweenValues) ^ bigowl4dq:hasOperator(?valCond2, bigowl4dq:EqualOrGreater) ^ bigowl4dq:rangeMax(?valCond, ?max) ^ bigowl4dq:valueConditionNumber(?valCond2, ?max) -> bigowl4dq:isCompleteCondition(?cond, true) In the case where the condition contains the default value (e.g., ‘‘- ’’), this means that the condition will meet any input value; therefore, this condition will be triggered by any input value; then, the condition is established as complete. Case 2: Completeness in a condition bigowl4dq:hasClause(?br, ?cl) ^ bigowl4dq:hasAssociatedHitPolicyCriteria(?cl, ?hit) ^ bigowl4dq:hasCondition(?cl, ?cond) ^ bigowl4dq:hitCondition(?cond, ?hit) ^ bigowl4dq:hasValueCondition(?cond, ?valCond) ^ bigowl4dq:valueConditionString(?valCond, ?str) ^ swrlb:equal(?str, "-") -> bigowl4dq:isCompleteCondition(?cond, true) - Rules for checking the completeness in a clause. One or more conditions form a clause. Two rules have been defined to assess the completeness of a clause. The first case is when a clause contains only one condition. In this case, the rule checks the completeness of the condition and inherits its value, as shown below. Case 1: Completeness in a clause bigowl4dq:hasCondition(?cl, ?cond) ^ bigowl4dq:hasClause(?br, ?cl) ^ bigowl4dq:isCompleteCondition(?cond, ?valCond) ^ swrlb:equal(?valCond, true) ^ bigowl4dq:numberCondition(?cl, ?num) ^ swrlb:equal(?num, 1) ^ bigowl4dq:hitCondition(?cond, ?hit) ^ bigowl4dq:hasAssociatedHitPolicyCriteria(?cl, ?hit) -> bigowl4dq:isCompleteClause(?cl, true) The second case is when a clause has two or more conditions; in this case, the rule assesses the completeness of each pair of conditions, and only when all its conditions meet the completeness, the clause does it. Information and Software Technology 167 (2024) 107378 16 C. Barba-González et al. [70] P.N. Mendes, H. Mühleisen, C. Bizer, Sieve: linked data quality assessment and fusion, in: Proceedings of the 2012 Joint EDBT/ICDT Workshops, 2012, pp. 116–123. [71] E. Rahm, H.H. Do, et al., Data cleaning: Problems and current approaches, IEEE Data Eng. Bull. 23 (4) (2000) 3–13. [72] D. Calvanese, M. Dumas, Ü. Laurson, F.M. Maggi, M. Montali, I. Teinemaa, Semantics and analysis of DMN decision tables, in: M. La Rosa, P. Loos, O. Pastor (Eds.), Business Process Management, Springer International Publishing, 2016, pp. 217–233. [73] J. Poore, T. Nemecek, Reducing food’s environmental impacts through producers and consumers, Science 360 (6392) (2018) 987–992. [74] Á. Valencia Parra, Analysis of big data architectures and pipelines: Challenges and opportunities, in: Máster Universitario en Ingeniería Informática, 2019, pp. 1–96. [75] J. de Haro-Olmo, Á. Valencia-Parra, Á.J. Varela-Vaca, J.A. Álvarez-Bermejo, M.T. Gómez-López, ELI: an IoT-aware big data pipeline with data curation and data quality, PeerJ Comput. Sci. 141 (2023) 113450. [76] C. Schader, M. Curran, A. Heidenreich, J. Landert, J. Blockeel, L. Baumgart, B. Ssebunya, S. Moakes, S. Marton, G. Lazzarini, U. Niggli, M. Stolze, Accounting for uncertainty in multi-criteria sustainability assessments at the farm level: Improving the robustness of the SMART-Farm Tool, Ecol. Indic. 106 (2019) 105503. [77] E. Jerhamre, C.J.C. Carlberg, V. van Zoest, Exploring the susceptibility of smart farming: Identified opportunities and challenges, Smart Agric. Technol. 2 (2022) 100026, http://dx.doi.org/10.1016/j.atech.2021.100026, URL https: //www.sciencedirect.com/science/article/pii/S2772375521000265. [78] C. Schader, L. Baumgart, J. Landert, A. Muller, B. Ssebunya, J. Blockeel, R. Weisshaidinger, R. Petrasek, D. Mészáros, S. Padel, C. Gerrard, L. Smith, T. Lindenthal, U. Niggli, M. Stolze, Using the sustainability monitoring and assessment routine (SMART) for the systematic analysis of trade-offs and synergies between sustainability dimensions and themes at farm level, Sustainability 8 (3) (2016) 1–20. [79] W.G. de Almeida, R.T. de Sousa, F.E. de Deus, G. Daniel Amvame Nze, F.L.L. de Mendonça, Taxonomy of data quality problems in multidimensional Data Warehouse models, in: 2013 8th Iberian Conference on Information Systems and Technologies (CISTI), 2013, pp. 1–7. [80] Á. Valencia Parra, Á.J. Varela Vaca, M.T. Gómez López, P. Ceravolo, CHAMALEON: framework to improve data wrangling with complex data, in: ICIS 2019: 4th International Conference on Information Systems (2019), Association for Information Systems (AIS), 2019, pp. 1–17. [81] A. Even, G. Shankaranarayanan, Utility-driven assessment of data quality, Data Base 38 (2) (2007) 75–93, http://dx.doi.org/10.1145/1240616.1240623. [82] M. Janssen, H. van der Voort, A. Wahyudi, Factors influencing big data decision-making quality, J. Bus. Res. 70 (2017) 338–345. [83] United States Department of Agriculture, Data from: A field-scale sensor network data set for monitoring and modeling the spatial and temporal variation of soil moisture in a dryland agricultural field, 2007, https://agris.fao.org/agrissearch/search.do?recordID=US2019X00214, Last accessed on 2021-06-02. [84] C.K. Gasch, D.J. Brown, C.S. Campbell, D.R. Cobos, E.S. Brooks, M. Chahal, M. Poggio, A field-scale sensor network data set for monitoring and modeling the spatial and temporal variation of soil water content in a dryland agricultural field, Water Resour. Res. 53 (12) (2017) 10878–10887, http://dx.doi.org/10.1002/2017WR021307,arXiv:https:// agupubs.onlinelibrary.wiley.com/doi/pdf/10.1002/2017WR021307, URL https: //agupubs.onlinelibrary.wiley.com/doi/abs/10.1002/2017WR021307. [85] F.J. de Haro-Olmo, Á. Valencia-Parra, Á.J. Varela-Vaca, J.A. Álvarez-Bermejo, Data curation in the Internet of Things: A decision model approach, Comput. Math. Methods (2021) e1191. [86] C.K. Gasch, D.J. Brown, E.S. Brooks, M. Yourek, M. Poggio, D.R. Cobos, C.S. Campbell, A pragmatic, automated approach for retroactive calibration of soil moisture sensors using a two-step, soil-specific correction, Comput. Electron. Agric. 137 (2017) 29–40, http://dx.doi.org/10.1016/j.compag.2017.03.018, URL https://www.sciencedirect.com/science/article/pii/S0168169916304288. [87] C. Wohlin, P. Runeson, M. Höst, M.C. Ohlsson, B. Regnell, Experimentation in Software Engineering, Springer, 2012, http://dx.doi.org/10.1007/978-3-642- 29044-2. [88] R.J. Wieringa, Design Science Methodology for Information Systems and Software Engineering, Springer, 2014. [89] A. Benítez-Hidalgo, C. Barba-González, J. García-Nieto, P. Gutiérrez-Moncayo, M. Paneque, A.J. Nebro, M. del Mar Roldán García, J.F.A. Montes, I.N. Delgado, TITAN: A knowledge-based platform for Big Data workflow management, Knowl.-Based Syst. 232 (2021) 107489, http://dx.doi.org/10.1016/j. knosys.2021.107489.