Different strategies, process-driven and data-driven, can improve data quality. Batini et al. [10] reported that individual strategies used different techniques. Nevertheless, each strategy is aimed at enhancing the quality of data.

Process-driven strategy

Process-driven strategy is aimed at redesigning all the processes, or processes and alters the data to enhance quality. This type of strategy has two additional techniques, namely, process redesign and process control. Process redesigns eliminate the causes of low data quality and thereby add a new process to generate high quality, while Process controls check and manage the manufacturing process [10]. Process-driven strategy per-forms better in the long period; it obliterates the causes of quality problems.

Data-driven strategy

A data-driven strategy is intended at improving the data quality through adjustment of the data value, nonstop. In this strategy, there are some related improvement methods of data-driven strategy, and they are data error correction and localization, normalization, standardization, source trustworthiness, cost optimization, and schematic integration [10]. Data-driven strategy is more expensive than process-driven, and it is effective in the short term.

Different strategies, process-driven and data-driven, can improve data quality. Batini et al. [10] reported that individual strategies used different techniques. Nevertheless, each strategy is aimed at enhancing the quality of data.

Data are the new oil; it can be stored, elaborated, and retrieved through a soft-ware-based process and can communicate through a network [11]. Researchers have classified data into different types in different are-as. Table 1 presents three different data types according to their classifications, implicitly or explicitly. Another classification of data is considering data as a product and is classified into three types as presented in Table 2.

Adapted from Sidi et al. [12].

Adapted from Sidi et al. [12].

Data can also be classified based on austerity to measure and achieve quality data, and this can be classified into two classes, namely, aggregated and elementary data. Data managed using active processes and that represent atomic occurrences of the world are referred to as elementary data, and they are age, sex, and others. Data gathered from elementary data and applied with an aggregate function is called aggregate data, tax, income, and others.

Data quality is a multidimensional idea as data quality is assessable by considering a range of scopes applicable to a specific field. There are three different approaches to assess the data quality dimensions: empirical, theoretical, and intuitive.

Empirical method

The empirical approach is employed to analyze data gathered from consumers; it is used to determine the dimensions they use to evaluate if the data are fit for use in their frameworks. The empirical method is a general term for research techniques that conclude from observable evidence [13].

Theoretical method

This approach focuses on data becoming incomplete during the built-up process. The theoretical approach can provide a complete set of data quality dimensions. This data quality approach assumes that an information system represents a real-time world system as the users see. The quality of data dimension is derived based on possible inconsistencies between how the user sees the practical system as inferred from the information system and the view collected by openly observing the practical system.

Intuitive method

The data quality dimension in an intuitive approach is chosen based on the researcher’s experience and knowledge about what features are vital in a specific framework. The merit of using this method is that it allows the individual study to choose the dimensions important to the specific goals of the framework.

Different processes obstruct data; most of these processes affect data quality to a certain extent. The quality of data deteriorates differently based on the situation they are found. There are thirteen kinds of processes that lead to data problems; Figure 1 presents some of the causes of data problems which are gathered into three advanced categories.

The left side of Figure 1 presents the processes in bringing data from the outside through manual means or using data integration methods and some interfaces. The introduction of incoming data will bring forth errors during the data loading, transformation, and extraction processes; these problems are further magnified through the traffic of high data volumes. The right side contains the processes that handle the data manipulation. Some procedures in data manipulation are routine, while some are due to database restructure, periodic, mass data updates, and some other ad hoc actions. However, in the implementation stage, some of these procedures lack the resources, time, and reliable metadata necessary to realize data quality effects.

The bottom part highlights the chain of processes that cause correct data to be incorrect over time, without any physical alteration applied to it. It simply means the data values are intact, but their accuracy drops. Such a situation arises when a real-time object described by the data changes, but the data gathering processes do not capture the change, turning the old data inaccurate and obsolete.

Sidi et al. [12] further classified data quality problems into two different groups: single-source and multi-source problems. Previous research also categorizes data quality into four different forms, as illustrated in Table 3.

There are three vital roles in the framework of data quality; they are data custodian, data producers, and data consumers [15] [16]. Data custodians are the individuals managing the resources for impending use. Data producers are individuals, sources, or groups that produce the data. Data consumers are the data users who have access to the collected data. These roles in the context of data quality are related to a task. Producers of data are associated with the production of the data process. Data custodians are saddled with the securing, maintenance, and storage of data and give access to consumers that want to employ data for a specific process. Data quality works gauge its quality from the consumers’ perception of the data. In some dispensation data, custodians and producers are often categorized as the same unit.

It is a known fact from previous literature that the performance of machine learning and deep learning model is bounded by the quality of data used for model training and validation [17] [18]. There have been continuous efforts to improve the quality of models, but little effort has been recorded toward refining the quality of data. One of the essential requirements before employing a specific

Adapted from Sidi et al. [12].

dataset is to understand the dataset and its compatibility with the intended research work. Failure to do so will result in erratic inferences and incorrect analytics. Evaluating the data quality over intelligently developed metrics and developing matching transformation operations to solve the quality gaps helps ease the effort of data scientists for iterative mending of the machine learning algorithm to enhance the performance of the model. Due to these facts, it is imperative to design an algorithm or tool to reduce the execution time for data preparations. The data quality toolkit is the systematic measure of data quality to develop machine learning models. Presented in Figure 2 is the general structure of the positional stand of the data quality toolkit as a part of the typical data science pipeline. Gupta et al. [19] reported that a data quality toolkit was developed to address four primary objectives:

1) Analyze data quality at step 0 of a data science development, such that the data scientist can make informed inferences at the later stages of model development.

2) Offer explanations for cases of low data quality by highlighting the region responsible for the low quality.

3) Suggest actions to enhance the data quality and provide a smooth way to implement the suggestions.

4) In the presence of an auto-generated report highlighting the operations history on the dataset for smooth reference to all the changes made to the original data.

The data toolkit can serve as a decision support system (DSS) to ensure robust decision-making processes. These decisions can involve selecting a model that identifies issues such as features dropping and noise irrelevant labels, determining the quality of data, and providing a response to the data acquisition procedure. The data quality toolkit contains a set of metrics that measure data problems serving as a pointer of the consumed data readiness for downstream machine learning tasks. The data quality dimensions are measured on a scale of 0 - 1, whereby a score specifies the absence of practical issues for the individual quality feature. The data quality toolkit comprises three key components of data remediation, data quality measurement, and data readiness report.

The data remediation mechanisms can be invoked to enhance and adjust the data quality issues acknowledged by the individual quality metrics. Based on the type of quality measurement, this step may include a human in the loop process to integrate human review and response before data transformation. After using all the data remediation functions on tasks, this step writes the updated data to the system file. Data remediation also helps to confirm whether the user wants to include a new class tag for data in updated data or not [19].

Data quality measurement is critical and remains the most significant component in the toolkit. A data quality measurement takes in the data and assesses the data quality based on different quality features. The quality scores measure quality measurement’s interest as an actual number between 1 and 0. There is no issue when a score of 1 is recorded in the ingested data for that metric. A textual explanation is provided to help users interpret the score meaning. Also, recommendations are offered to protect the user’s technique that can be applied to fix the issues acknowledged.

The data readiness report is a shareable asset that is the one-stop-shop of the standard quality of the input data and a record of operations and remediation carried out [20]. It is fundamentally a single source for understanding the data quality and the series of transformations that has been undertaken. Over time data scientists spend much time tackling and exploring different data quality issues. The data readiness report is a complete report of all data properties and quality issues, such as the lineage of the data that illustrates how the data evolve. Furthermore, a data readiness report is more of a dynamic and evolving document that can be generated at any time during the interaction with the toolkit.

Data quality monitoring is a framework that incessantly controls the data quality present in an information system through metric reports or periodically conducting data profiling. Data quality measurement is different from data quality assessment; the base is to evaluate a single data quality score from a range of observations. Data quality can be monitored through an application that monitors various information systems over some time. Presented in Figure 3 is the structure of the data quality monitoring application; it consists of four main components, namely, 1) Metadata gathering and computation of data quality assessment results; 2) data quality depository; 3) Time-series data analysis; and 4) a user-friendly interface to monitor and track the data quality development of the observed system.

Data quality dimension is a characteristic of information used for organizing information and data requirements. Over time, the most frequently mentioned data quality dimensions are completeness, accuracy, consistency, and timeliness, with different definitions from researchers. Wang and Strong [22] proposed different categories of data quality dimensions. Figure 4 presents six categories of data quality dimensions, namely, accuracy, completeness, timeliness, constituency, uniqueness, and validity.

Completeness

Completeness can be defined as the degree to which data are of sufficient depth, breadth, and scope for the operation at hand [23].

Accuracy

Accuracy is the degree to which data are error-free, reliable, and accurate [23]. It can be further defined as the closeness between a value v and v’, deliberated as the correct representation of the real-time occurrence that v represents [11].

Timeliness

Timeliness shows how to present data for the job at hand. It is inspired by the fact that it is possible to have current data that are ineffective because they are late for a specific usage [11]. It can be further defined as the degree to which data are satisfactorily updated for the operation at hand [22]. Timeliness is defined as the datedness associated with volatility and age; the age of information or an object is the measure of how old that piece of information or object is [24]. Volatility is a measure of information unpredictability with the incidence of change of the value for an entity characteristic.

Consistency

Consistency is about how data is achieved in a system with specified restrictions. Consistency, as addressed in a database system, is seen as an integrity restriction, and it is highly regarded as crucial data quality problem.

Different proposals report the classification of data quality dimensions instead of conservative data quality aspects [26]. The most widely considered proposals for data quality dimensions are highlighted below:

Proposal by Nauman

In the proposal of Nauman [27], data quality dimensions are defined as specific to incorporated web information systems where quality is considered an accumulated value of numerous data quality standards. Under this proposal, four different categories of data quality dimensions play a vital role in web information systems. These categories are intellectual, content-related, instantiation-related criteria, and technical. From the mentioned criteria, some criteria will be vital for a particular application domain. The two domains are stock information systems and search engines.

Proposal by Wang and Strong

In this proposal, data quality is determined by cross-examining data consumers. The data quality category for Wang and Strong proposal was intrinsic, accessibility, contextual, and representational data quality [22]. This proposal was used for many data quality studies to develop diverse data quality models and structures in different contexts.

Proposal by Bovee et al.

Proposal by Bovee et al. [24] considers data quality dimensions by examining the view of data consumers and developed a theoretical model consisting of four different features. This model was adapted from Wang and Strong [22], and the features have been compared with this proposal to assess data quality with all its essential dimensions or attributes that determine the quality in any field; this proposal has four different attributes, namely, relevance, integrity, interpretability, and accessibility [24]. The integrity attribute is classified into four sub-attributes completeness, existence, consistency, and accuracy.

Decision-making in an institution or organization is centered on information and data retrieved through data analysis that offers different insights to generate an accurate and dependable process. The way data is essential in some institutions is the same way its quality must be checked. Sidi et al. [12] focus on a review of data quality dimensions to be employed for a proposed framework that combines statistical methods and data mining to evaluate dependencies present between dimensions; this study further illustrates how retrieving knowledge can increase process quality.

Previous and existing studies agree that data quality is a cyclic process that must be continuously carried out. However, most data quality tools permit the assessment of data sources only at a certain point in time. Therefore, it is imperative to schedule and automate data by the user to support an excellent data evaluation. The quality of data monitoring ensures the assessment of applied data quality developments and compares various system states. Ehrlinger and Wöß [21] developed a data quality monitoring tool and investigated correct data quality metrics for continuous monitoring and to enable the development of the consistent technique of storing data quality assessment results.

Nowadays, companies and institutions have invested a lot in data, and most of their decisions are guided by data collected from the field or sales. Therefore, inaccurate data information will ultimately compromise the decision process. It is essential to verify the source, integrity, and quality of their data to ensure smooth decision making. Schelter et al. [28] proposed a system for automating the processes in verifying data quality. The verification system provides a declarative application program interface that merges quality limitations with user-defined authentication code and permits unit tests for data. The system further leverages the machine learning algorithms and incremental data quality validation on increasing datasets.

The validation of the quality of data is essential in present-day data-enabled applications. Unexpected behaviors in downstream services like machine learning models and production pipelines are recorded due to errors in the founding data. Over the years, unforeseen data quality problems have been resolved through manual and monotonous restoring processes volatilely. Growing usage of a large dataset must be ingested into non-relational stores such as data lakes. It is even worse when the features of the to-be ingested dataset change periodically, and field expertise is not available to define the data quality limitations. Redyuk et al. [29] proposed a data-centric technique that automates data quality assessment in scenarios painted earlier on. Here, the proposed technique does not require field expert to define limitations and rules; the technique is built such that the data self-adapt to periodic and temporal changes in the data features. The technique was improved and evaluated on five real-time datasets having both artificial and real generated errors.

Günther et al. [30] reported that not all organizations check data quality before any decision-making processes. Therefore, a simplified methodology was proposed to execute data quality evaluations and enhance the understandability of its results. This work was aimed at making data quality usable and accessible to small and medium-sized enterprises. The technique uses context-related semi-structured and structured data as the input and employs a range of generic test criteria applicable to various fields and tasks. The methodology employed is assessed using data gathered from the manufacturing execution system and enterprise resource planning.

The growing importance of big data has introduced new challenges to data quality. Accordingly, various research studies have sought to identity the quality dimensions of big data. Table 4 presents an overview of the big data quality classification.

The existing literature on various dimensions of the data quality management has been discussed in this section. The following section will briefly explain the methodology used in identifying the literature and criteria used for selecting specific papers for inclusion in this review.

Assessment of data quality is a process of evaluating collected data to determine whether the program’s objectives are met, and the data is the right type that has the quality to support the intended objectives. The results obtained from the data quality assessment will determine the relevancy, accessibility, completeness, accuracy, currency, and consistency of the data. Data quality is highly reliant on the underlying data reporting and management systems. To generate quality data, functional components must be available in all levels, ranging from data sources, gathering, encoding, transfer, data checking, reviewing, and processing. The data quality assessment tools are intended to 1) evaluate the sources of the data; 2) validate the collected data quality; 3) design strategic actions to enhance the system and data. The data quality assessment intends to evaluate the quality of gathered data, thereby aiming to a) ensure that the information retrieved reflects the authenticity on the field and has all the six properties of data quality dimension; b) assess the capacity of the dataset system by managing, capturing, reporting, and processing quality data; and lastly c) to design and implement measures to strengthen data gathering and management at all levels.

Components of data quality assessment

The components of data quality assessment are divided into three main groups, namely, a) agreement with data conversion and submission criteria; b) data authentication; and c) system evaluations.

Component 1: Agreement with data conversion and submission criteria: Providing an excellent database that offers complete and appropriate data to partners and management is an important objective that enhances an informed and better decision-making process. This first component deals with the timelessness and completeness of the submitted data. Data reporting standards and requirements are pre-set and provided at all levels.

Component 2: Data authentication: In ensuring data authentication, templates, forms, and documents are employed to capture activities available in the database. Data verification helps confirm the consistency and accuracy of the data from different sources it was gathered from through cross-checking the reported data and information. This component will detect, track, and resolve discrepancies and errors in the database.

Component 3: System evaluations: Data quality is highly reliant on the systems put in place. Systems with solid backgrounds produce improved data quality. The system evaluation will address the following facts:

1) Abilities of the staff involved in gathering and management of data.

2) Technical aid and capacity building

3) Data gathering, management, and processing involve using patterns and paper-based forms, conversion and submission, data quality mechanism, reporting and application, recovery, and storage.

The data quality assessment is intended to help aid the management or researchers to understand the fundamental limitations and problems faced during data gathering, management, and processing, also to determine possible sources of error in data gathering, detect measures to enhance the capabilities of users involved in the process and reinforce data management at all levels.

To ensure the continued quality of data over time, improving data quality dimensions through an efficient process is vital. Owing to this reason, numerous researchers have worked to propose a model and methodologies for organized data quality management. Wang [42] proposed a model named Total Data Quality Management (TDQM). TDQM was intended to support the idea of data being a product. This idea will ensure the production of high data quality through duplication of the physical production of a high-quality product. The TDQM extends the total quality management framework that employs physical production. The approaches of TDQM start with a description of the production information. The information production (IP) has its features and requirements to achieve a high-quality data state. The second approach is the quality of information metrics developed and employed to measure the information product. The result obtained from the measurement is analyzed and used the pattern recognition, Pareto chart, and statistical process control. Time data quality management was developed to manage the quality of data present in the database, and present technologies involving big data analytics may limit the practice.

Another data quality management model is the Aim Quality model by Lee et al. [43], which comprises of service and product performance model for information quality. This tool measures information quality and analyses the gap within the information quality to enhance the quality of information retrieved. The medium of information retrieval and evaluation in this method is a questionnaire. Further inferences are unraveled through statistical analysis, which involves identifying information quality problems.