Ensuring data quality in machine learning (ML) systems has become increasingly complex as regulatory requirements expand. In the European Union (EU), frameworks such as the General Data Protection Regulation (GDPR) and the Artificial Intelligence Act (AI Act) articulate data quality requirements that closely parallel technical concerns in ML practice, while also extending to legal obligations related to accountability, risk management, and human rights protection. This paper presents a qualitative interview study with EU-based data practitioners working on ML systems in regulated contexts. Through semi-structured interviews, we investigate how practitioners interpret regulatory-aligned data quality, the challenges they encounter, and the supports they identify as necessary. Our findings reveal persistent gaps between legal principles and engineering workflows, fragmentation across data pipelines, limitations of existing tools, unclear responsibility boundaries between technical and legal teams, and a tendency toward reactive, audit-driven quality practices. We also identify practitioners' needs for compliance-aware tooling, clearer governance structures, and cultural shifts toward proactive data governance.

Understanding how data quality aligns with regulatory requirements in machine learning (ML) systems presents a critical challenge for practitioners navigating the evolving EU regulatory landscape. To address this, we first propose a practical framework aligning established data quality dimensions with specific EU regulatory requirements. Second, we conducted a comprehensive online survey with over 180 EU-based data practitioners, investigating their approaches, key challenges, and unmet needs when ensuring data quality in ML systems that align with regulatory requirements. Our findings highlight crucial gaps between current practices and regulatory expectations, underscoring practitioners' need for more integrated data quality tools and better collaboration between technical and legal practitioners. These insights inform recommendations for bridging technical expertise and regulatory compliance, ultimately fostering responsible and trustworthy ML deployments.

Data errors are widespread in real-world databases and severely impact downstream applications, such as machine learning pipelines or business analytics reports. Causes of such errors are manifold and can arise during both the design phase and the operational phase of a database. Some error types, such as missing values, duplicate tuples, or constraint violations, are widely recognized; others, such as disguised missing values or word transpositions, remain underexplored. Existing attempts to define and classify errors in data offer valuable but limited taxonomies, mostly informal and not covering the full range of error types. With the rise of AI, practitioners must increasingly detect and correct statistical errors such as bias and outliers, which are rarely considered within existing error taxonomies. This catalog presents a comprehensive list of 35 distinct error types, including both data errors (e.g., missing values, duplicate tuples) and error indicators (e.g., outliers, bias) for tabular data, classified into three non-overlapping categories: missing, incorrect, and redundant. For each error type, we provide a formal definition and practical example, and resolve terminological inconsistencies across related work. Our catalog enables researchers and practitioners to address various error types and systematically implement error-specific detection and cleaning strategies in data quality tools.

Modern artificial intelligence (AI) applications require large quantities of training and test data. This need creates critical challenges not only concerning the availability of such data, but also regarding its quality. For example, incomplete, erroneous, or inappropriate training data can lead to unreliable models that produce ultimately poor decisions. Trustworthy AI applications require high-quality training and test data along many quality dimensions, such as accuracy, completeness, and consistency.We explore empirically the relationship between six data quality dimensions and the performance of 19 popular machine learning algorithms covering the tasks of classification, regression, and clustering, with the goal of explaining their performance in terms of data quality. Our experiments distinguish three scenarios based on the AI pipeline steps that were fed with polluted data: polluted training data, test data, or both. We conclude the paper with an extensive discussion of our observations.

This editorial summarizes the content of the Special Issue on Data quality dimensions in Data FAIRification design and processes of the Journal of Data and Information Quality (JDIQ).

Data quality comprises a large set of dimensions, each describing a specific aspect. The assessment of these dimensions requires the collection of simple statistics, the identification of syntactic problems and factual errors, as well as organizational and business aspects to be considered. With the current trend in data-oriented sciences and the increasing reliance on machine learning methods and AI systems, the challenges of poor data quality are ever more apparent. Even recent legislation, such as the European AI Act, mentions data quality requirements for training data; with it, the notion of data quality extends to novel dimensions, such as fairness, diversity, or explainability. We present a modular approach to the assessment of data quality, based on five key facets: data, source, system, task, and human.

Data-oriented applications, their users, and even the law require data of high quality. Research has divided the rather vague notion of data quality into various dimensions, such as accuracy, consistency, and reputation. To achieve the goal of high data quality, many tools and techniques exist to clean and otherwise improve data. Yet, systematic research on actually assessing data quality in its dimensions is largely absent, and with it, the ability to gauge the success of any data cleaning effort. We propose five facets as ingredients to assess data quality: data, source, system, task, and human. Tapping each facet for data quality assessment poses its own challenges. We show how overcoming these challenges helps data quality assessment for those data quality dimensions mentioned in Europe's AI Act. Our work concludes with a proposal for a comprehensive data quality assessment framework.

Artificial intelligence systems depend critically on high-quality data, yet real-world datasets are often imperfect. Label noise, such as incorrect or biased labels, can lead to suboptimal model decisions. While label noise has garnered increasing attention, existing research primarily examines random noise, employs simpler models, or relies on limited evaluation criteria. To address this, we introduce FAULT LINES, a comprehensive, model-agnostic benchmark comprising 15 datasets systematically corrupted with diverse types of label noise, paired with an evaluation framework. This resource supports the evaluation of data cleaning pipelines and guides the design of models that are robust, in both performance and fairness, to label noise. We benchmark the robustness to label noise of 22 state-of-the-art classification models, including gradient boosting, transformers, and fairness-oriented models. Our findings show that many models maintain strong performance under high random noise (e.g., up to 40% noise leads to only a modest reduction in Robust GBDT performance). However, these models are significantly less robust to even small amounts of biased noise (<10%), which can cause substantial performance drops (e.g., 7% noise reduces ResNet's AUC by 4.4% on average) or maintain apparent stability at the expense of severe fairness degradation (e.g., MLP's Predictive Parity difference increases by 700% under 30% biased noise in the ACS Unemployment dataset). We investigate how different model architectures handle the impact of biased noise. Notably, transformer-based models appear more robust than boosting models when handling biased noise, though this advantage depends on tuning and comes with higher variance. Finally, we identify key factors for ML practitioners to mitigate the effects of label noise, including model selection, dataset analysis, and preprocessing.

Data quality is crucial in machine learning (ML) applications, as errors in the data can significantly impact the prediction accuracy of the underlying ML model. Therefore, data cleaning is an integral component of any ML pipeline. However, in practical scenarios, data cleaning incurs significant costs, as it often involves domain experts for configuring and executing the cleaning process. Thus, efficient resource allocation during data cleaning can enhance ML prediction accuracy while controlling expenses. This paper presents COMET, a system designed to optimize data cleaning efforts for ML tasks. COMET gives step-by-step recommendations on which feature to clean next, maximizing the efficiency of data cleaning under resource constraints. We evaluated COMET across various datasets, ML algorithms, and data error types, demonstrating its robustness and adaptability. Our results show that COMET consistently outperforms feature importance-based, random, and another well-known cleaning method, achieving up to 52 and on average 5 percentage points higher ML prediction accuracy than the proposed baselines.

Data-oriented applications, their users, and even the law require data of high quality. Research has divided the rather vague notion of data quality into various dimensions, such as accuracy, consistency, and reputation. To achieve the goal of high data quality, many tools and techniques exist to clean and otherwise improve data. Yet, systematic research on actually assessing data quality in its dimensions is largely absent, and with it, the ability to gauge the success of any data cleaning effort. We propose five facets as ingredients to assess data quality: data, source, system, task, and human. Tapping each facet for data quality assessment poses its own challenges. We show how overcoming these challenges helps data quality assessment for those data quality dimensions mentioned in Europe's AI Act. Our work concludes with a proposal for a comprehensive data quality assessment framework.

Artificial Intelligence (AI) supported job interviewing, i.e., one-sided automated applicant interviews assessed by AI-based systems, presents itself as a new mainstream solution in hiring, promising to be more efficient and effective than human recruiters, but also fairer and more objective. Selecting this technology as an illustrative case, we focus on a central element in the development of fair AI: the issue of (training) data quality (DQ). ML models with unsuitable, biased, or erroneous training data is a major source of bias in AI-based applications and therefore potentially discriminatory, unfair outcomes. However, DQ is often cast aside as one of many technical factors contributing to the overall quality of ML-based systems; this approach runs the risk of understating its crucial relevance. We select salient issues along the technology lifecycle to take a detailed look at the interrelation of fairness and DQ, illustrating how both fairness and DQ must be understood in a broad sense, taking into account normative considerations beyond technical aspects, to facilitate desirable outcomes such as the promotion of diversity, the prevention of discrimination, and the protection of workers’ rights.

Modern artificial intelligence (AI) applications require large quanti-ties of training and test data. This need creates critical challenges not only concerning the availability of such data, but also regarding its quality. For example, incomplete, erroneous or inappropriate training data can lead to unreliable models that produce ultimately poor decisions. Trustworthy AI applications require high-quality training and test data along many dimensions, such as accuracy, completeness, consistency, and uniformity. We explore empirically the correlation between six of the tra-ditional data quality dimensions and the performance of fifteen widely used ML algorithms covering the tasks of classification, regression, and clustering, with the goal of explaining ML results in terms of data quality. Our experiments distinguish three scenarios based on the AI pipeline steps that were fed with polluted data: polluted training data, test data, or both. We conclude the paper with an extensive discussion of our observations and recommendations, alongside open questions and future directions to be explored.

Column headers are among the most relevant types of meta-data for relational tables, because they provide meaning and context in which the data is to be interpreted. Headers play an important role in many data integration, exploration, and cleaning scenarios, such as schema matching, knowledge base augmentation, and similarity search. Unfortunately, in many cases column headers are missing, because they were never defined properly, are meaningless, or have been lost during data extraction, transmission, or storage. For example, around one third of the tables on the Web have missing headers. Missing headers leave abundant tabular data shrouded and inaccessible to many data-driven applications.We introduce a fully automated, multi-phase system that discovers table column headers for cases where headers are missing, meaningless, or unrepresentative for the column values. It leverages existing table headers from web tables to suggest human-understandable, representative, and consistent headers for any target table. We evaluate our system on tables extracted from Wikipedia. Overall, 60% of the automatically discovered table headers are exact and complete. Considering more header candidates, top-5 for example, increases this percentage to 72%.

The research area of data profiling consists of a large set of methods and processes to examine a given dataset and determine metadata about it. Typically, different data profiling tasks address different kinds of metadata, comprising either various statistics about individual columns (Single-column Analysis) or relationships among them (Dependency Discovery). Among the basic statistics about a column are data type, header, the number of unique values (the column's cardinality), maximum and minimum values, the number of null values, and the value distribution. Dependencies involve, for instance, functional dependencies (FDs), inclusion dependencies (INDs), and their approximate versions. Data profiling has a wide range of conventional use cases, namely data exploration, cleansing, and integration. The produced metadata is also useful for database management and schema reverse engineering. Data profiling has also more novel use cases, such as big data analytics. The generated metadata describes the structure of the data at hand, how to import it, what it is about, and how much of it there is. Thus, data profiling can be considered as an important preparatory task for many data analysis and mining scenarios to assess which data might be useful and to reveal and understand a new dataset's characteristics. In this thesis, the main focus is on the single-column analysis class of data profiling tasks. We study the impact and the extraction of three of the most important metadata about a column, namely the cardinality, the header, and the number of null values. First, we present a detailed experimental study of twelve cardinality estimation algorithms. We classify the algorithms and analyze their efficiency, scaling far beyond the original experiments and testing theoretical guarantees. Our results highlight their trade-offs and point out the possibility to create a parallel or a distributed version of these algorithms to cope with the growing size of modern datasets. Then, we present a fully automated, multi-phase system to discover human-understandable, representative, and consistent headers for a target table in cases where headers are missing, meaningless, or unrepresentative for the column values. Our evaluation on Wikipedia tables shows that 60% of the automatically discovered schemata are exact and complete. Considering more schema candidates, top-5 for example, increases this percentage to 72%. Finally, we formally and experimentally show the ghost and fake FDs phenomenon caused by FD discovery over datasets with missing values. We propose two efficient scores, probabilistic and likelihood-based, for estimating the genuineness of a discovered FD. Our extensive set of experiments on real-world and semi-synthetic datasets show the effectiveness and efficiency of these scores.

Inclusion dependencies are an important type of metadata in relational databases, because they indicate foreign key relationships and serve a variety of data management tasks, such as data linkage, query optimization, and data integration. The discovery of inclusion dependencies is, therefore, a well-studied problem and has been addressed by many algorithms. Each of these discovery algorithms follows its own strategy with certain strengths and weaknesses, which makes it difficult for data scientists to choose the optimal algorithm for a given profiling task. This paper summarizes the different state-of-the-art discovery approaches and discusses their commonalities. For evaluation purposes, we carefully re-implemented the thirteen most popular discovery algorithms and discuss their individual properties. Our extensive evaluation on several real-world and synthetic datasets shows the unbiased performance of the different discovery approaches and, hence, provides a guideline on when and where each approach works best. Comparing the different runtimes and scalability graphs, we identify the best approaches for certain situations and demonstrate where certain algorithms fail.

This article is an extended abstract of our work published at VLDB’2018. The full paper is available at www.vldb.org/pvldb/vol11/p880-berti-equille.pdf . Functional dependencies (FDs) play an important role in maintaining data quality in relational databases. They can be used to enforce data consistency and guide data repairs. In this work, we investigate the problem of missing values and its impact on FD discovery. When using exist- ing FD discovery algorithms, some genuine FDs could not be detected precisely due to missing values and some non-genuine FDs can be discovered even though they are caused by missing values depending on the considered semantics for NULL values. We define the notion of gen- uineness of FDs and propose algorithms to compute the FD genuineness score. This can be used to identify genuine FDs among the set of all valid dependencies that hold on the data. We evaluate the quality of our method over various real-world and semi-synthetic datasets with extensive experiments. The results show that our method performs well for relatively large FD sets and is able to accurately capture genuine FDs.

Functional dependencies (FDs) play an important role in maintaining data quality. They can be used to enforce data consistency and to guide repairs over a database. In this work, we investigate the problem of missing values and its impact on FD discovery. When using existing FD discovery algorithms, some genuine FDs could not be detected precisely due to missing values or some non-genuine FDs can be discovered even though they are caused by missing values with a certain NULL semantics. We define a notion of genuineness and propose algorithms to compute the genuineness score of a discovered FD. This can be used to identify the genuine FDs among the set of all valid dependencies that hold on the data. We evaluate the quality of our method over various real-world and semi-synthetic datasets with extensive experiments. The results show that our method performs well for relatively large FD sets and is able to accurately capture genuine FDs.

Data preparation and data profiling comprise many both basic and complex tasks to analyze a dataset at hand and extract metadata, such as data distributions, key candidates, and functional dependencies. Among the most important types of metadata is the number of distinct values in a column, also known as the zeroth-frequency moment. Cardinality estimation itself has been an active research topic in the past decades due to its many applications. The aim of this paper is to review the literature of cardinality estimation and to present a detailed experimental study of twelve algorithms, scaling far beyond the original experiments. First, we outline and classify approaches to solve the problem of cardinality estimation - we describe their main idea, error-guarantees, advantages, and disadvantages. Our experimental survey then compares the performance all twelve cardinality estimation algorithms. We evaluate the algorithms' accuracy, runtime, and memory consumption using synthetic and real-world datasets. Our results show that different algorithms excel in different in categories, and we highlight their trade-offs.