Unfair Inequality in Education: A Benchmark for AI-Fairness Research.

Skip to main content An official website of the United States government Here's how you know Here's how you know Official websites use .gov A .gov website belongs to an official government organization in the United States. Secure .gov websites use HTTPS A lock ( Lock Locked padlock icon ) or https:// means you've safely connected to the .gov website. Share sensitive information only on official, secure websites. Search Log in Dashboard Publications Account settings Log out Search… Search NCBI Primary site navigation Search Logged in as: Dashboard Publications Account settings Log in Search PMC Full-Text Archive Search in PMC Journal List User Guide PERMALINK Copy As a library, NLM provides access to scientific literature. Inclusion in an NLM database does not imply endorsement of, or agreement with, the contents by NLM or the National Institutes of Health. Learn more: PMC Disclaimer | PMC Copyright Notice Sci Data . 2026 Mar 2;13:572. doi: 10.1038/s41597-026-06827-x Search in PMC Search in PubMed View in NLM Catalog Add to search Unfair Inequality in Education: A Benchmark for AI-Fairness Research Joseph Giovanelli Joseph Giovanelli 1 University of Bologna, DISI, Viale Risorgimento 2, Bologna, 40123 Italy Find articles by Joseph Giovanelli 1, ✉ , Matteo Magnini Matteo Magnini 1 University of Bologna, DISI, Viale Risorgimento 2, Bologna, 40123 Italy Find articles by Matteo Magnini 1 , Giovanni Ciatto Giovanni Ciatto 1 University of Bologna, DISI, Viale Risorgimento 2, Bologna, 40123 Italy Find articles by Giovanni Ciatto 1 , Angel S Marrero Angel S Marrero 2 University of La Laguna, CEDESOG, San Cristóbal de La Laguna, S/C de Tenerife, 38200 Spain Find articles by Angel S Marrero 2, ✉ , Andrea Borghesi Andrea Borghesi 1 University of Bologna, DISI, Viale Risorgimento 2, Bologna, 40123 Italy 3 Alma Mater Research Center for Human-Centered Artificial Intelligence, Bologna, 40121 Italy Find articles by Andrea Borghesi 1, 3 , Gustavo A Marrero Gustavo A Marrero 2 University of La Laguna, CEDESOG, San Cristóbal de La Laguna, S/C de Tenerife, 38200 Spain Find articles by Gustavo A Marrero 2 , Roberta Calegari Roberta Calegari 1 University of Bologna, DISI, Viale Risorgimento 2, Bologna, 40123 Italy Find articles by Roberta Calegari 1 Author information Article notes Copyright and License information 1 University of Bologna, DISI, Viale Risorgimento 2, Bologna, 40123 Italy 2 University of La Laguna, CEDESOG, San Cristóbal de La Laguna, S/C de Tenerife, 38200 Spain 3 Alma Mater Research Center for Human-Centered Artificial Intelligence, Bologna, 40121 Italy ✉ Corresponding author. Received 2025 Jun 11; Accepted 2026 Feb 4; Collection date 2026. © The Author(s) 2026 Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by-nc-nd/4.0/ . PMC Copyright notice PMCID: PMC13066447  PMID: 41771897 Abstract This paper introduces a novel benchmark dataset designed to support fairness-oriented research in artificial intelligence within the educational domain. The dataset originates from longitudinal survey data collected by the Agencia Canaria de Calidad Universitaria y Evaluación Educativa, encompassing comprehensive information from students, families, and teachers across the Canary Islands, Spain. It includes detailed student profiles and academic trajectories, covering multiple years of academic performance outcomes. The original data is characterised by a high-dimensional and sparse feature space, which presents challenges for direct application in AI workflows. To address these challenges while minimising the risk of introducing bias during preprocessing, we provide a curated version of the dataset specifically tailored for AI applications. This version preserves the statistical properties of the original data and is accompanied by detailed documentation of the preprocessing steps, including strategies for dimensionality reduction and fairness preservation. The dataset is intended as a resource for the research community, enabling studies on fairness, predictive modeling, and educational analytics. We describe its structure, content, and preparation process. Subject terms: Education, Education Background & Summary Artificial Intelligence (AI) fairness has gained an increasing amount of relevance in the development of intelligent systems, particularly those that learn patterns and gain insights from data 1 . Data-driven systems are inherently vulnerable to inheriting and amplifying biases embedded in their training data, potentially resulting in discriminatory or unjust outcomes. Marginalised or disadvantaged groups—whether defined by gender, ethnicity, age, religion, political affiliation, or other sensitive attributes—can be disproportionately affected, thereby reinforcing pre-existing social inequalities and obstructing upward mobility. While high-dimensionality and sparsity are common in domains such as finance, educational data presents unique challenges. Unlike financial indicators, which are often objective and continuous, educational features are predominantly contextual and subjective, derived from multiple stakeholders (e.g., the ’triangulation’ of student, family, and teacher perspectives). Furthermore, the ’missingness’ in education is frequently tied to socio-economic vulnerabilities; for example, parents from marginalised backgrounds may intentionally withhold information to avoid perceived stigma—a distinct form of MNAR data that requires domain-specific ethical handling rather than standard statistical imputation. Finally, educational datasets combine heterogeneous sources—student performance, family socio-economic indicators, teacher practices, and school-level context—often collected through subjective questionnaires. This results in a feature space that is not only large but also semantically diverse, with categorical and ordinal variables intertwined with sensitive information. To address these concerns, the research community has made significant progress in developing techniques for detecting, measuring, and mitigating algorithmic bias. A key element in this process is the use of benchmark datasets, which serve as standardised tools for evaluating fairness-aware algorithms and uncovering systemic biases in AI models. Widely used benchmarks include UCI’s Adult dataset 2 , the German Credit dataset 3 , and the COMPAS dataset on recidivism risk prediction 4 . Several datasets address fairness in education, but each has key limitations. FAIR-Edu 5 focuses on faculty, not students. Eedi 6 and EdNet 7 provide large-scale interaction logs but lack demographic or contextual data. The NELS dataset 8 includes rich socio-demographic variables but is rarely used in AI due to access and format issues. MOOC datasets 9 offer insight into online learning but reflect a digitally privileged population. Overall, most datasets lack demographic depth, contextual richness, and accessibility for effective fairness analysis in education. Fairness benchmarks are scarce 10 , 11 . Existing datasets are often constrained in terms of domain coverage, demographic diversity, and contextual relevance. As such, there is a growing need for new, high-quality benchmarks that support a broader range of fairness-related tasks and foster the development of equitable AI systems across diverse real-world applications. Additionally, datasets intended for fairness-related studies often require processing and manipulation to simplify downstream tasks 12 . This step is delicate, as it can inadvertently introduce or amplify biases—such as when proxy variables for sensitive attributes are retained or when normalization methods disproportionately impact specific subgroups 13 . This study presents a new dataset 14 explicitly collected to enable fairness research, combining a variety of data sources that provide an extensive overview of demographical information. We also demonstrate how the data was preprocessed in order not to ensure the introduction of any form of additional bias - excluding the bias potentially present in the original data. We provide both the original data and the preprocessed data for the sake of maximum transparency. The proposed dataset comes from the high-stakes domain of education, with a particular aim towards ensuring equitable access and opportunities for all students. It includes students from Las Palmas de Gran Canaria and Santa Cruz de Tenerife in the Canary Islands, Spain. This dataset goes beyond basic student information by incorporating detailed data on family background, teacher profiles, and academic performance. The original data primarily consists of a longitudinal survey, where responses were collected according to EU privacy regulations via questionnaires and carefully anonymised. We then present a refined version that has been carefully preprocessed to facilitate the training of AI algorithms. The preprocessing phase is conducted using co-design methodologies 15 , involving key stakeholders and domain experts from both the education and economics sectors. This participatory approach supports the identification of sensitive features, the validation of techniques for handling missing values, and the definition of criteria for feature selection and transformation. A co-design methodology is very important in this context, to guarantee that technical decisions reflect the domain-specific knowledge, ethical considerations, and fairness objectives defined by stakeholders directly impacted by educational policies and outcomes. Our proposed dataset comes with several benefits. First, it focuses on a contemporary educational setting, hence ensuring the data captures the nuances of modern educational environments in a real-world scenario. Second, the dataset is very rich: not only in terms of size but also in terms of dimensionality and plurality of perspectives. It encompasses student background, family dynamics, and teacher perspectives, allowing for a more nuanced study of the factors impacting students’ success. Both the raw and preprocessed datasets and their metadata are made available to the research community. For the sake of traceability and reproducibility, we store them on Zenodo ( https://zenodo.org/records/17592007 ); we also provide the code used to create the data and process it ( https://github.com/aequitas-aod/Unfair-Inequality-Education ). We highlight how our dataset differs from a selection of related datasets in Table 1 . The table highlights how no existing benchmark jointly provides accessibility, demographic depth, and fairness-aware preprocessing. We also notice how NELS 16 is methodologically rich but operationally inaccessible for most AI researchers. Table 1. Comparison of educational datasets along key dimensions relevant to AI fairness research. Columns correspond to datasets, rows to dataset characteristics. Characteristic Proposed Benchmark NELS (1988) 16 Eedi 37 EdNet 7 MOOC datasets 38 Educational domain Public primary and secondary education Secondary education (U.S.) Online formative assessment Intelligent tutoring systems Online higher education Demographic attributes Rich, multi-source Rich None None Limited, self-reported Socio-economic context Extensive (families, schools) Extensive None None Limited Teacher and school information Yes Partial No No No Longitudinal structure Yes (multi-year) Yes Limited Yes (interaction logs) Yes Population representativeness Public education census/sampling Nationally representative Task-specific sample Platform-specific users Self-selected, digitally privileged Fairness-oriented design Explicitly designed for fairness No No No No Bias-aware preprocessing Yes (validated bias preservation) No No No No Missing data semantics Explicit (MCAR/MAR/MNAR) Implicit Not documented Not documented Not documented Accessibility for AI research Public, AI-ready Restricted; legacy formats Public Public Public Primary limitations Geographic scope Access barriers; outdated context; not formatted for modern AI No demographics; semantic sparsity (hard to understand the motivation of the students’ performance) No socio-economic context for fairness; lacks sensitive attributes Strong selection bias (represents a digitally-privileged population); limited contextual depth Open in a new tab In the remainder of the paper, we use the term fairness-preserving preprocessing to denote preprocessing strategies that aim to avoid introducing additional unfairness into the data. In this context, fairness preservation is operationalised as distributional parity preservation, i.e., the maintenance of the conditional distributions between sensitive attributes and target variables before and after preprocessing. When referring to this property, we occasionally use the term bias preservation to indicate that preprocessing does not amplify or distort existing biases already present in the raw data. Methods The Agencia Canaria de Calidad Universitaria y Evaluación Educativa (ACCUEE) ( https://archive.is/zOSkB ) is a public institution in the Canary Islands, which aims to monitor the quality of education services in the region. To serve this purpose, it collects data concerning students, curricula, and schools. Concerning students, the collected data includes information on their academic performance in Mathematics, Spanish Language (local native language), and English, plus questionnaire responses from students, families, teachers, and school principals, which attempt to capture the socio-economic and cultural background of the students, as well as the situations of their schools. The data is collected from students in the 3 rd and 6 th grades of primary education, and in the 4 th grade of secondary education. The collection process was repeated over four academic years, from 2015–16 to 2018–19. Figure 1 illustrates the dataset construction process that is described in more detail in the following subsections. Fig. 1. Open in a new tab Overview of the methodology employed to generate the dataset; the aggregation method applied to process the questionnaires is displayed as well – the aggregation method is always the mean value. Data Collection The raw data collected by ACCUEE comes in tabular form. The table comes with 83,857 rows and 554 columns. Each row refers to a single student at a given grade and academic year. However, the total amount of involved students is lower, as information for each student is collected multiple times: students in 3 rd grade (primary school) during A.Y. 2015–16 are sampled again in their 6 th grade. Primary education data for the A.Y. 2015–16 and 2018–19 are gathered through a comprehensive census over the entire population. For other grades and academic years, the data is collected through sampling. Longitudinal data is also included: students in 3 rd grade (primary school) during A.Y. 2015–16 are sampled again in their 6 th grade. The columns of the table represent relevant features collected for each student. Overall, the 555 columns represent features from six categories – each category of answers corresponds to a distinct questionnaire: identifiers (8 columns): various sorts of identifiers involving the student (e.g., at the school or ACCUEE level), and their academic situation (school, grade, academic year); performance features (6 columns), containing the student’s scores in Mathematics, Spanish Language, and English; students’ answers (154 columns) to a questionnaire concerning their own experiences with the school system (including but not limited to their access to resources, their relationship with teachers, their satisfaction level, etc.); principals’ answers (138 columns) to a questionnaire concerning the school the student is enrolled in; families’ answers (91 columns) to a questionnaire concerning the socio-economic conditions of the student; teachers’ answers (158 columns) to a questionnaire concerning their workload, satisfaction, and methodology. Due to the longitudinal nature of the data collection process, the number of rows in the dataset does not correspond to the number of unique students. Each row represents a student observed at a specific grade and academic year, and individual students may therefore appear multiple times across different waves of data collection. Overall, the dataset comprises 83,857 rows (corresponding to 40,391 students), with repeated measurements occurring when the same student is sampled at different educational stages. In particular, students enrolled in the 3rd grade of primary education during the 2015-16 academic year were followed up and resampled in their 6th grade, while other grades were sampled cross-sectionally; the number of students involved in the longitudinal analysis is 17,521. To improve clarity, Table 2 summarizes the sampling timeline, explicitly indicating which grades were surveyed in each academic year and highlighting the longitudinal overlaps between cohorts. Table 2. Education data by grade and school year. 2015–2016 2016–2017 2017–2018 2018–2019 7D23 grade 20430 3202 3290 7D26 grade 20711 3237 3301 19855 Secondary education 4 grade 3132 3291 3408 Open in a new tab Data Pre-processing Questionnaires were standardised; consequently, the features obtained from them are either categorical or numerical. However, the data is not clean as missing values are present, and they are not evenly distributed across the features. Given the high dimensionality of the dataset, preprocessing is necessary to make it more manageable and suitable for the tasks we intend to address. Missing Values According to 17 , any preprocessing strategy dealing with missing data should keep the nature of the missing data into account. The dataset contains various types of missing data. Following the categorisation of 18 : Missing Completely at Random (MCAR) values are unrelated to any observed or unobserved cause. In our scenario, missing data can arise if participants fail to complete their questionnaires, for instance, due to sporadic absences from school or work. However, because of their randomness, we consider MCAR values as unrelated concerning other features in the dataset. Missing at Random (MAR) values are related to observed causes, but they are un related w.r.t. other missing values. In our scenario, this is often caused by certain questions not being administered in specific grades or schools. Hence, no assumption should be made in imputing MAR values as there is no causality with the fact that they are missing. Missing Not at Random (MNAR) (Missing Not At Random) values occur when missing data is linked to unobserved factors or other missing values. In our context, this often happens when participants intentionally withhold sensitive information—perhaps due to feelings of shame or fear of judgment. For example, families in disadvantaged situations may avoid completing questionnaires or refuse to disclose details about their living conditions because of time constraints or reluctance. Since these omissions may be influenced by underlying social or psychological factors, it is important to consider the mechanisms behind missingness. People might hide information to prevent discrimination, and ignoring this possibility can lead to biased results. It is important to notice that explicitly recognising MNAR values highlights the ethical dimension of data preprocessing. Imputing or modeling such data without understanding why it is missing risks introducing bias and compromising fairness in subsequent analyses. The choice to avoid automated imputation (e.g., Multiple Imputation by Chained Equations - MICE) for MNAR features hinges on the specific requirements of a fairness benchmark. Recent literature indicates that standard imputation techniques can inadvertently reduce the variance of protected subgroups or introduce “imputation bias” that masks existing inequalities 19 . Since the missingness in our dataset is often informative (e.g., reflecting socio-economic barriers), we prioritise data fidelity over data completeness 18 . This ensures that the benchmark reflects the real-world “sparsity” that AI systems must navigate in educational settings, preventing the creation of a “synthetic” fairness that does not hold in practice 20 . Notice that these categories are defined at the value level: a dataset’s column may contain missing values of different sorts. Hence, determining the ‘missingness behaviour’ of each column is non-trivial. For this reason, while labelling our dataset’s features as MCAR, MAR, or MNAR, we leveraged domain experts and missing values analysis 17 to be as accurate as possible. In the educational domain, MNAR values are not merely a technical hurdle but a proxy for social exclusion. Our methodology distinguishes itself from generic AI preprocessing by treating these gaps as sensitive signals. While a financial model might impute a missing stock price to maintain continuity, an educational AI must recognise that a missing response regarding ’home resources’ may itself be a critical indicator of disadvantage that should not be ’smoothed over’ through automated imputation. The categorisation above is important because it impacts how missing values should be handled. Principles in 21 outline how random missing data (MCAR and MAR) do not affect the overall distribution of the observed data, hence they can be safely ignored 17 ; ‘ignorability’ here means handling missing values as ‘ordinary’ observed values (i.e., without imputing them explicitly). Notice that the same word is used with different semantics in other contexts (e.g., 22 ). We treat MCAR and MAR features by aggregating them with other variables, following the same dimensionality reduction strategies discussed below. Conversely, MNAR values are deemed “non-ignorable” 21 because the missing values could be intentionally withheld – most commonly, this happens with features carrying sensitive information, and handling them requires designing ad-hoc imputation or augmentation strategies 23 which are inherently domain- and goal-specific, and come with ethical implications. Hence, for MNAR features, we choose not to apply any transformation that could compromise the integrity or fairness of the dataset, hence prioritising transparency, and leaving users free to decide how to handle them in their use cases. Accordingly, we simply mark sensitive features as ‘MNAR’ via ad-hoc metadata. We categorized the features as MCAR/MAR/MNAR. If a feature was MNAR, it was because it was sensitive, and therefore, we did not apply any processing in order not to alter its distribution. If, instead, it was MCAR or MAR, we applied the transformations without excessive concern, since the missing value did not conceal an unexpressed intention (such as shame about one’s economic status). For example, aggregation operations using the mean combine features related to the same topic but collected at different points (e.g., frequency of checks, relationship with the teacher), and for each row, compute the average of the available values. There is certainly some distortion, but it is minimal; to verify this, we performed additional tests on the similarity of the distributions before and after the processing. Fairness-aware Dimensionality Reduction Our raw dataset is highly affected by the curse of dimensionality 24 , mostly due to the data coming from standardised questionnaires with many detailed questions. To address this issue, we implement a series of preprocessing steps to create compact features reflecting various aspects of students’ lives. As part of our methodology, responses from the four questionnaires—students, principals, families, and teachers—are semantically grouped and aggregated by thematic area. Specifically, we organise questions that share similar structures, answer types, and topical relevance into coherent groups. This approach allows us to derive a set of aggregated features, which we categorise into three distinct types: degree of agreement to a certain statement (e.g., questions such as “The teacher is fair / helps us / listens to us” are aggregated into a single score: ‘extent of teacher affinity’), frequency of a certain activity (e.g., questions such as “How often do you use the computer at home/school/library?” are aggregated into a single score: ‘frequency of computer usage’), holding (or not) a certain characteristic (e.g., questions such as “Did you repeat your 1 st / 2 nd /… grade” are aggregated into a single score: ‘has repeated’). To estimate the aforementioned features, we apply the following transformation pipeline: Feature Selection: we discard features with duplicated information or with low information content. Specifically, we evaluate candidate features by checking at their semantics, and examining their Pearson correlation 25 with the performance features (i.e., students’ scores in school), or the predictive importance as determined by the GINI index 26 . Feature Creation . We employ of two techniques: binary encoding: boolean responses are converted into binary representations, allowing for the creation of categorical features capturing nuanced variations in responses; this approach preserves the information contained in the original questionnaire data while facilitating further analysis and modelling. mean aggregation: feature values are aggregated through means, summarising responses on similar topics, hence creating more interpretable and manageable features for subsequent analyses. We justify the use of mean aggregation for handling MCAR and MAR values within thematic groups. Unlike predictive imputation, which estimates a value based on other features, mean aggregation leverages the internal consistency of the survey instrument 18 . In educational psychology, when multiple items assess the same latent construct (e.g., “academic expectations”), the mean of available responses provides a more robust and less biased estimate than cross-feature prediction, particularly in high-dimensional, sparse environments where model-based imputation is prone to overfitting 27 – 29 . Normalisation: We normalise score distributions to ensure consistency and comparability across different types of responses. Here, we leverage Min-Max normalisation. Min-max aggregation was employed for its simplicity: we wanted to use an aggregation method that would preserve the original distribution and that generate an output distributed in the range [0-1], to have a more direct comparison across all features (to facilitate follow-up research studies). The dataset resulting from the preprocessing stage retains a significant portion of the data, with 80,174 out of 83,857 rows preserved, while significantly reducing the number of columns to 139 features, derived from the original 541 questionnaire responses. This represents a reduction of approximately 73.9% in the overall dataset dimensionality. Besides, the preprocessed dataset exhibits a 33.1% rate of missing values, whereas the rate of the original data is 45.1%. Unlike other domains, sparsity here is not merely technical but socially driven: missing values often correspond to sensitive socio-economic questions, making aggressive imputation risky for fairness. This ethical dimension amplifies the complexity of preprocessing educational data compared to domains like finance. Privacy and Ethical Concerns ACCUEE follows strict protocols for data collection and informed consent in its diagnostic evaluations. These procedures ensure participant confidentiality and full compliance with current data protection laws. Students take standardised tests to assess key competencies. Alongside, both students and school staff complete questionnaires with contextual information about the home environment, the school, and the teaching process. Data is collected through structured paper or digital forms during the school year. This information is used to analyse outcomes and support efforts to improve educational quality and equity. Importantly, test results have no academic impact and are not recorded in student transcripts. All responses are anonymous and confidential. Participants’ data is protected by two main safeguards. First, all collected information is covered by statistical secrecy: it can only be used for statistical analysis, and everyone involved—whether individuals or institutions—must maintain confidentiality even after their role ends (as required by Article 22 of Law 1/1991 on Statistics of the Autonomous Community of the Canary Islands). Second, ACCUEE fully complies with data protection regulations, including Spain’s Organic Law 3/2018 on Personal Data Protection and Digital Rights, and the EU’s General Data Protection Regulation (GDPR). Personal data is treated as strictly confidential, and informed consent is mandatory for participants or their legal guardians, in line with GDPR Articles 13 and 14. To safeguard identities, data is pseudo-anonymised: each participant is assigned a code instead of using names. Additionally, the municipality and school of each student are anonymised, preventing identification based on geographic or institutional details. Teachers and professionals with access to personal or family information are also bound by a confidentiality obligation. While schools may collect necessary student data for educational purposes, such as family background or academic progress, this information cannot be used for other purposes without explicit consent. In short, ACCUEE ensures all data collection and processing respects the privacy and rights of participants, fully aligned with legal requirements. Ethical Checklist IRB committee details The data were collected by the Canarian Agency for Quality Assessment and Accreditation (Agencia Canaria de Calidad Universitaria y Evaluación Educativa - ACCUEE) as part of its legally mandated diagnostic educational evaluations. These procedures are strictly regulated under Spanish Law 1/1991 (Statistical Law of the Canary Islands), Organic Law 3/2018 on Personal Data Protection, and the EU General Data Protection Regulation (GDPR). As ACCUEE operates under public mandate with established legal and ethical frameworks, no additional Institutional Review Board (IRB) approval number applies. Participants recruitment Participants (students and school staff) were included through the official diagnostic evaluations conducted by ACCUEE across schools in the Canary Islands. Participation is part of the educational evaluation process, which is designed to improve educational quality and equity. Details on participants’ consent Participants, or their legal guardians in the case of minors, consented to participation in the standardized tests and contextual questionnaires, as well as to the collection and use of their data for research and statistical analysis purposes, in accordance with GDPR Articles 13 and 14. Consent does not extend to open data sharing; data is used exclusively for statistical and research purposes under strict confidentiality. How informed consent was obtained Informed consent was obtained from participants or their legal guardians (for minors) following GDPR requirements. The process included clear information about the purpose of the evaluations, the voluntary nature of participation, and data protection safeguards. Steps taken to protect participant information All data are pseudo-anonymized: participants are assigned codes instead of names, and identifiers such as municipality and school are anonymized. The information is subject to statistical secrecy under Article 22 of Law 1/1991, which prohibits use for purposes other than statistical analysis. Data are also protected under Organic Law 3/2018 and the GDPR. De-anonymisation risk The dataset has been fully anonymized and pseud-anonymised in compliance with GDPR standards, and no geographic or institutional identifiers remain that could allow re-identification. Consequently, the risk of de-anonymisation is negligible. Data Records This dataset 14 proposes a novel benchmark specifically designed for AI fairness research in education. It can be used for challenging tasks aimed at improving students’ performance and reducing dropout rates, which are also discussed in the paper to emphasize significant research directions. By prioritising fairness, this benchmark aims to foster the development of fairness-preserving AI solutions, promoting equal educational access and outcomes for all students. The data record can be found at the following URL: https://zenodo.org/records/17592007 , with the following document identifier 10.5281/zenodo.11171862. The structure of the data record 14 is the following: BENCHMARK contains: the proposed dataset ( dataset.csv ), the mask for dealing with missing values ( missing_mask.csv ), and the meta-columns providing grouping criteria and sample weights for each student ( meta_cols.csv ). RAW_DATA includes: the original dataset ( original.csv ), and the intermediate stages of the pre-processing and validation pipeline ( split , pre_processed , and validation ). RES contains the documentation, including: the transformation mapping each column of the original dataset to the proposed one, along with the missingness category, and original text ( meta_data_mapping.csv ), the value type, the domains of each column of the proposed datasets, and the correlation with the original column group ( meta_data_stats.json ), and the statistical indices of the validation pipeline ( bias_preservation_results.json ). SRC contains the source code for running the pre-processing and corresponding analysis: pre_processing and stats contain the code for the two corresponding tasks, and pre_processing.py and split.py are two entry points. Finally, Dockerfile and requirements.txt set up the environment for running the applications across multiple platforms and with Python, respectively. Technical Validation In this section, we aim to validate the preprocessing stage, ensuring that the essential information and statistical properties of the original dataset 𝒟 original = { X original , S , Y } are retained in the new preprocessed dataset 𝒟 preproc = { X preproc , S , Y } , where X original are the original questionnaire answers, X preproc are the new derived features, and S and Y are the sensitive and target features respectively. We performed two different analyses: (i) measuring feature-by-feature the mutual information between the two datasets, and (ii) comparing the estimated bias w.r.t. each sensitive feature between the two datasets. The first analysis is a vertical assessment that computes the correlation between the new aggregated features of X i ∈ X preproc and their corresponding raw input groups from X j , ⋯ ∈ X original . Our objective is to demonstrate that the preprocessing methodology captures and preserves the semantic content of the original features. The second analysis is a horizontal assessment, verifying that the relationship between sensitive features S and target features Y remains consistent across the datasets. To this end, we evaluate whether the preprocessed dataset preserves the fairness of outcomes estimated using the raw dataset, specifically showing that any potential bias concerning sensitive groups is not amplified during preprocessing. Features’ Mutual Information To assess whether the information content of the preprocessed features was preserved for the original dataset, we conducted a feature-wise comparison using the Hirschfeld-Gebelein-Rényi (HGR) coefficient 30 . The HGR coefficient is a non-linear dependence measure defined as the supremum of the Pearson correlation between two random variables A and B under arbitrary non-linear transformations: HGR ( A , B ) = sup f , r ∈ ℱ cov ( f ( A ) , r ( B ) ) σ ( f ( A ) ) ⋅ σ ( r ( B ) ) , where f and r range over all square-integrable functions in the Hilbert space ℱ . The HGR is a powerful tool for capturing a wide range of statistical dependencies beyond linear correlations. This makes it particularly well-suited to our goal, as it allows us to determine whether the preprocessed features retain the full spectrum of dependency structures (linear and non-linear) present in the original data. By computing the HGR between each original feature and its transformed counterpart, we can quantitatively assess the extent to which the preprocessing technique preserves the intrinsic informational relationships of the original data distribution. While the exact HGR is theoretically incomputable, we adopted two estimation algorithms implemented in our evaluation framework to approximate this measure. Double Kernel Estimator This approach follows the method introduced by Giuliani et al . 31 , extending it for HGR estimation. It maps the input variables into polynomial kernel spaces. For two feature vectors a and b , the copula transformations are represented as F a ⋅ α and G b ⋅ β , where F a = [ a , a 2 , …, a h ] and R b = [ b , b 2 , …, b k ] are kernel expansions, and α , β are mixing coefficients learned through global optimization. The objective is to find the non-linear projections that maximize the correlation between the transformed representations. Single Kernel Estimator A more computationally efficient variant, the single-kernel estimator tests directional dependencies by correlating either a transformed input against the raw target or vice versa. It solves an unconstrained least-squares problem to estimate α and β for transformations f ( a ) and g ( b ), again based on polynomial kernels. While it does not fully explore the symmetric HGR formulation, it provides a gradient-friendly and fast-to-compute approximation, suitable for large datasets. We applied both estimators to all aggregated features X i preproc ∈ X preproc , comparing them to the corresponding group of original features X j origin … ∈ X origin from which they were derived. The HGR values computed via both methods were consistently high and closely aligned. The average HGR coefficient across all feature pairs was μ = 99.78 with a standard deviation σ = 0.72, indicating an almost perfect dependency structure between original and preprocessed variables. These results confirm that the dimensionality reduction process preserved the underlying semantics of the data without introducing significant distortions. Fairness-Preservation We aim to ensure that the preprocessing step does not alter the relationship between the input features and either the target or sensitive features. Specifically, we assess whether the derived students’ features X preproc , in the new preprocessed dataset D preproc = { X preproc , S , Y } , maintain the same conditional behaviour with respect to each sensitive feature S i ∈ S (e.g., gender) and target feature Y j ∈ Y (e.g., Math’s score) as in the original dataset D original = { X original , S , Y } . To do so, for each target feature Y j , we train an estimator h on X original and an estimator g on X preproc . Then, for each sensitive feature S i , we compare the resulting conditional distributions Y ^ i , j o r i g i n a l = P ( h ( X original ) ∣ S i ) and Y ^ i , j p r e p r o c = P ( g ( X preproc ) ∣ S i ) . All estimators are Random Forests with 500 trees, chosen for their ability to handle missing values and complex patterns. To ensure robustness and representativeness, we also validated our findings with XGBoost classifiers. In each case, we compare the conditional distributions using multiple statistical tools: the T-test, Kolmogorov-Smirnov (KS) test, and Chi-square test – along with two scalar indices: Maximum Difference (Max Diff) and Mean Absolute Error (MAE). Table 3 shows the results of the estimated difference between the distributions Stat , we aggregated this value by averaging the performance across the different target features Y j (i.e., Maths, English, and Spanish scores). Besides, we report the maximum reference value critic for not rejecting the null hypothesis (i.e., claiming that the new distributions have no significant changes). Follow the discussion for each test, along with the final p-value. In the following, we analyse the tests’ results with the corresponding p-values, as well as the indicator performance. Table 3. Statistical tests results in comparing the distributions P ( h ( X original )∣ S ) and P ( h ( X prep )∣ S ) according to Sensitive Features S . The differences between the distributions (stat) are averaged among the possible target features (math, Spanish, and English levels). The stat value, the lower the better, is colored red depending on how far it is from the critical value (critic). Sensitive Feature T-Test KS-Test Chi-Test Max Diff MAE Stat Critic Stat Critic Stat Critic Stat Critic Stat Critic s_gender red!01 e -13(±9 e -14) 2 red!270.13(±0.00) 0.5 red!332.677D1(±1.01) 8 red!40.04(±0.01) 1 red!20.02(±0.01) 1 s_has_repeated red!06 e -13(±4 e -12) 2 red!180.09(±0.03) 0.5 red!60.517D1(±0.11) 8 red!00.00(±0.00) 1 red!00.00(±0.00) 1 s_birth_country red!02 e -14(±2 e -14) 2 red!230.02(±0.00) 0.1 red!2881.677D1(±17.05) 292 red!60.06(±0.00) 1 red!10.01(±0.00) 1 s_nationality_country red!02 e -14(±2 e -14) 2 red!270.03(±0.01) 0.1 red!35109.33(±12.65) 308 red!90.09(±0.01) 1 red!10.01(±0.00) 1 p_island red!04 e -14(±5 e -14) 2 red!180.07(±0.04) 0.38 red!152.357D1(±2.38) 16 red!60.06(±0.06) 1 red!10.01(±0.01) 1 p_public_or_private red!02 e -12(±2 e -12) 2 red!220.11(±0.03) 0.5 red!231.807D1(±1.72) 8 red!00.00(±0.00) 1 red!00.00(±0.00) 1 f_mother_education_level red!06 e -13(±4 e -13) 2 red!150.04(±0.00) 0.27 red!62.097D1(±0.26) 36 red!00.00(±0.00) 1 red!00.00(±0.00) 1 f_father_education_level red!07 e -13(±2 e -13) 2 red!150.04(±0.00) 0.27 red!51.877D1(±0.19) 36 red!00.00(±0.00) 1 red!00.00(±0.00) 1 f_mother_employment_status red!02 e -13(±1 e − 12) 2 red!190.07(±0.00) 0.35 red!10.217D1(±0.06) 20 red!00.00(±0.00) 1 red!00.00(±0.00) 1 f_father_employment_status red!04 e -13(±6 e -13) 2 red!130.04(±0.02) 0.35 red!10.247D1(±0.04) 20 red!00.00(±0.00) 1 red!00.00(±0.00) 1 f_mother_place_of_birth red!03 e -14(±8 e -14) 2 red!140.05(±0.02) 0.38 red!10.157D1(±0.05) 16 red!40.04(±0.01) 1 red!20.02(±0.00) 1 f_father_place_of_birth red!02 e -14(±2 e -14) 2 red!210.08(±0.00) 0.38 red!10.177D1(±0.04) 16 red!50.05(±0.00) 1 red!20.02(±0.00) 1 f_student_place_of_birth red!04 e -14(±6 e -14) 2 red!110.04(±0.00) 0.38 red!10.217D1(±0.13) 16 red!50.05(±0.01) 1 red!20.02(±0.00) 1 f_extent_of_books_at_home red!03 e -13(±2 e -12) 2 red!160.06(±0.02) 0.35 red!40.717D1(±0.03) 20 red!00.00(±0.00) 1 red!00.00(±0.00) 1 f_parental_education_expectations red!02 e -14(±2 e -14) 2 red!140.05(±0.01) 0.33 red!30.617D1(±0.05) 24 red!160.16(±0.02) 1 red!40.04(±0.00) 1 f_mother_occupation red!01 e -12(±2 e -12) 2 red!180.07(±0.02) 0.38 red!30.417D1(±0.02) 16 red!00.00(±0.00) 1 red!00.00(±0.00) 1 f_father_occupation red!01 e -12(±2 e -12) 2 red!140.05(±0.02) 0.38 red!30.497D1(±0.04) 16 red!00.00(±0.00) 1 red!00.00(±0.00) 1 f_monthly_household_income red!02 e -14(±8 e -15) 2 red!160.04(±0.01) 0.26 red!31.407D1(±0.41) 40 red!210.21(±0.04) 1 red!20.02(±0.00) 1 f_ESCS red!06 e -14(±1 e -12) 2 red!170.06(±0.00) 0.33 red!71.597D1(±0.08) 24 red!00.00(±0.00) 1 red!00.00(±0.00) 1 Open in a new tab T-Test Analysis The independent two-sample T-test evaluates whether the means between the distributions Y ^ i , j o r i g i n a l and Y ^ i , j p r e p r o c differ significantly. The test computes the statistic by dividing the mean difference between distributions by the standard error of the difference: t = μ original − μ preproc δ 2 original n original + δ 2 preproc n preproc where μ and δ 2 are group means and variances, and n is sample size. The critical value is derived from the T-distribution table at 95% confidence ( α = 0.05). In Table 3 , all Stat values are orders of magnitude smaller than the Critic value, indicating no significant mean differences. This holds across all sensitive features, confirming that preprocessing retains statistical properties with a p-value = 1 for all cases. KS-Test The Kolmogorov-Smirnov test compares the cumulative distribution functions (CDFs) of the two distributions Y ^ i , j o r i g i n a l and Y ^ i , j p r e p r o c , leveraging the empirical CDF estimation F ( ⋅ ). For each sample y ^ k ∈ Y ^ i , j , this simply counts y ^ ≤ y ^ k ∀ y ^ ∈ Y ^ i , j . Then, it calculates the maximum difference: D = sup k ∣ F orig ( y ^ k ) − F prep ( y ^ k ) ∣ , where F orig and F preproc are the empirical CDF computed on Y ^ i , j o r i g i n a l and Y ^ i , j p r e p r o c respectively, and sup is the supremum of the set of distances. The critical value depends on the combined sample size and is computed as c ( α ) n orig + n prep n orig ⋅ n prep , where c ( α ) is a constant determined by the desired significance level α . As shown in Table 3 , the KS statistics have no case exceeding the critical threshold, even though we observe higher values compared to other tests. This is due to the nature of the test, which is sensitive to any kind of difference in distribution shape, including location, scale, and skewness. However, the p-value is always equal to 1, except for the target feature level_MAT and the sensitive feature s_nationality_country, in which the p-value is 0.98. Chi-square Test The Pearson’s chi-square test evaluates whether the frequency distributions of categorical outcomes differ between Y ^ i , j o r i g i n a l and Y ^ i , j p r e p r o c across sensitive groups. For each target-sensitive feature pair ( y ∈ Y , s ∈ S ), we construct contingency tables C o r i g i n = Count ( Y ^ i , j o r i g i n = y ∩ S = s ) and C p r e p r o c = Count ( Y ^ i , j p r e p r o c = y ∩ S = s ) . The test statistic is computed by comparing the contingency tables under the null hypothesis of identical distributions: χ 2 = ( O p r e p r o c − O o r i g i n a l ) 2 O o r i g i n a l . Again, the critical value depends on degrees of freedom (| Y | − 1)(| S | − 1). Compared to T/KS tests, the test exhibits higher Stat values for some cases (e.g.,s_birth_country, s_nationality_country) due to: (i) the test’s sensitivity to sample size–scaling with n orig ⋅ n prep /( n orig + n prep ), and (ii) the categorical nature of targets amplifies small relative differences. Nevertheless, all p-values are equal to 1, confirming that distributional parity (our operational notion of fairness preservation) is preserved between the datasets. Max Diff and MAE To fo further validate distribuibutional parity between Y ^ i , j o r i g i n a l and Y ^ i , j p r e p r o c , we compute two complementary metrics: To provide intuitive, non-parametric metrics of difference, we compute the maximum absolute difference (Max Diff) and mean absolute error (MAE) across all ( y ∈ Y , s ∈ S ) combinations: Max Diff = max y , s P ( h ( X original ) = y ∣ S i = s ) − P ( g ( X preproc ) = y ∣ S i = s ) , MAE = 1 ∣ Y j ∣ ∣ S i ∣ ∑ y , s P ( h ( X original ) = y ∣ S i = s ) − P ( g ( X preproc ) = y ∣ S i = s ) . Max Diff is the largest absolute frequency difference across the observed pairs ( y , s ). MAE is computed as the average absolute difference across all pairs. These values provide a scalar snapshot of how much distributions differ. The former remains below 0.1 for all the sensitive, except for f_parental_education_expectations (0.16 ± 0.02) and f_monthly_household_income (0.21 ± 0.04); while MAE remains below 0.05 in all cases. We believe these are isolated fluctuations. Max Diff is more sensitive to local peaks in the distribution, which may not reflect broader discrepancies as MAE, which values remain low for the considered cases. Usage Notes The proposed dataset 14 is intended to support both research and innovation, addressing a variety of use cases at the intersection of AI and education. In this section, we outline a set of key goals that can be pursued using the dataset, and we discuss the corresponding AI tasks required to achieve them. Fair selection of the most promising students for a given academic path ( G1 ) Selection processes strongly impact education systems. Ideally, all students would have the opportunity to pursue their preferred academic paths, but practical constraints—such as limited availability of placements—often mandate selection. In such scenarios, it is essential to ensure that the selection process does not reinforce or exacerbate existing biases. From a technical standpoint, achieving goal G1 requires the integration of two core AI tasks: a ranking task ( T1 ) and a supervised learning task ( T2 ). Task T2 involves predicting students’ academic performance – where fairness constraints must be explicitly considered – and can take the form of either classification or regression, depending on whether performance is treated as a categorical or continuous variable. Task T1 , in turn, consists of sorting students based on these predictions, potentially incorporating prediction confidence into the ranking process. Recognising the presence of both undersampled and oversampled schools in the dataset is crucial to ensure fairness across the tasks. To account for this, we provide sampling weights for the subset of the data collected through stratified or non-uniform sampling. These weights enable corrective interventions – such as data rebalancing – thus supporting more fair and representative analyses throughout the AI pipeline. Explainable underperforming students detection ( G2 ) To be effective, interventions aimed at improving student performance must be precisely targeted. This requires not only identifying whether a student is (potentially) underperforming, but also understanding the underlying reasons. Technically, goal G2 involves the following components: • detecting a performance threshold (task T3 ), • predicting a student’s performance (task T2 ), • comparing the two. Moreover, the model used for task T2 must be locally explainable 32 , allowing for the interpretation of individual predictions and the identification of contributing factors. For this purpose, the LIME method 33 can be employed to generate local explanations, helping stakeholders understand why a specific student is flagged as underperforming. A non-trivial challenge in this context is determining the performance threshold ( T3 ). This threshold could be defined a priori (e.g., based on established educational benchmarks) or inferred in a data-driven manner, such as through per-class, per-grade, or per-school distributions. In the latter case, special attention must be paid to prevent the thresholding process from perpetuating or introducing bias. Missing values are also highly relevant to this goal. In particular, there may be a non-random relationship between missingness and certain student circumstances—what is known as MNAR. We strongly encourage the use of the provided metadata to support appropriate handling of missing data, as well as the application of advanced techniques such as missing pattern analysis 23 , which can be used to augment the dataset and uncover informative absence patterns. Early drop-off prediction ( G3 ) Any intervention aimed at preventing student drop-off must be timely to be effective. Therefore, predicting drop-off in advance is a required capability to enable proactive measures. A rationale similar to that discussed for goal G2 applies here, with the additional complexity of forecasting an event that inherently results in data loss. From a technical standpoint, the key challenge for goal G3 lies in addressing sampling bias. By definition, students who drop out are no longer present in the educational system, and as a result, their data is unavailable in subsequent waves of the survey. This creates a structural absence – data about drop-offs is missing by design – and predictive models must therefore infer drop-off risk from incomplete information. To support this goal, an unsupervised feature engineering task ( T4 ) is necessary. Specifically, due to the longitudinal nature of the dataset, each student may appear at most twice. A lower number of recorded instances may be indicative of a drop-off, but this is not a sufficient condition. The key technical difficulty is distinguishing true drop-offs from students whose absence from the dataset is unrelated to dropout (e.g., due to temporary absence during data collection or changes in school enrollment). This task requires careful preprocessing and the design of robust heuristics or probabilistic models to estimate drop-off likelihood based on partial data. As with the previous goals, addressing missingness and sampling variability is critical to avoid introducing bias into the predictive process. Studying the impact of socio-economic factors on students’ performance ( G4 ) Understanding the impact of socio-economic factors on student performance is essential not only at the individual level but also for broader, systemic analyses. Extracting generalizable patterns from the dataset may reveal underlying structural inequalities and inform the design of more effective and equitable educational policies. Technically, goal G4 entails deriving a general model from the data, in which the contribution of each feature to performance-related outcomes is both observable and quantifiable. This can be approached using either globally explainable models – such as rule-based or decision tree models 34 – or probabilistic graphical models, such as Bayesian networks 35 . In the case of explainable models, the process involves training a performance predictor (task T2 ) and extracting global explanations from it (task T5 ), thereby highlighting the most influential features across the population. In the case of graphical models, the task consists of identifying a relevant set of variables—typically, socio-economic indicators from the preprocessed dataset—and learning their conditional dependencies directly from the data (task T6 ). Most-likely explanations of students’ performance ( G5 ) Gaining insight into the most plausible causes behind a student’s performance can enable more precise and individualized interventions, often without requiring the student to disclose sensitive personal information. Similar in spirit to goal G2 , goal G5 focuses on identifying the most likely explanation for an observed performance outcome by uncovering the complex interplay of influencing factors. This goal builds directly upon graphical models (cf. task T6 ) and can be addressed using the most probable explanation algorithm 36 , which allows inference of the most likely configuration of hidden or explanatory variables, given the observed data. Such approaches provide a principled and interpretable means to support decision-making processes in educational settings. Acknowledgements The work has been partially supported by the AEQUITAS project funded by the European Union’s Horizon Europe Programme (Grant Agreement No. 101070363). Author contributions J.G. led all phases of the creation and publication of the datasets, and he conducted the validation of the dataset. M.M. contributed to the processing of the original data, its cleaning, and wrote sections of the paper. G.C. contributed to the organization of the processed data and the writing of the paper by writing several sections and revising others. A.S.M. also contributed to the initial processing of the collected data and to the anonymization of the sensitive data. A.B. curated the publication of the dataset and of the code and contributed to the writing of the paper, while revisiting the manuscript as well. G.A.M. coordinated the data collection campaign as a liaison with all actors involved; he also contributed to the organization of the data. R.C. is the leader of the project, conceived the idea of creating and publishing the datasets, directed and coordinated the activities, provided critical revisions and the final approval of the manuscript, and operated as the main liaison between all partners involved. Data availability The data 14 is publicly available on Zenodo https://zenodo.org/records/17592007 ), under the licence Creative Commons Attribution 4.0 International (CC BY 4.0). The structure of the dataset is described in detail in Section; the dataset itself is auto-explanatory, as it contains a README file with all information required to understand how it is organized. Code availability The code used to pre-process the collected data is made available at the following webpage: https://github.com/aequitas-aod/Unfair-Inequality-Education . The code is made publicly available under the Apache License, Version 2.0 (January 2004). The main difference with respect to the data repository hosted on Zenodo (see Sec. is that in this latter repository, we also provide the code used to validate the data (the results of the validation are described in Sec.). Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Contributor Information Joseph Giovanelli, Email: [email protected]. Angel S. Marrero, Email: [email protected] References 1. Mehrabi, N., Morstatter, F., Saxena, N., Lerman, K. & Galstyan, A. A survey on bias and fairness in machine learning. ACM Comput. Surv. 54 , 115:1–115:35, 10.1145/3457607 (2022). [ Google Scholar ] 2. Becker, B. & Kohavi, R. Adult (a.k.a. “Census Income”) Dataset. UCI Machine Learning Repository (1996). 10.24432/C5XW20. 3. Hofmann, H. Statlog (a.k.a. “German Credit”) Dataset. UCI Machine Learning Repository (1994). 10.24432/C5NC77. 4. Propublica. Correctional offender management profiling for alternative sanctions (COMPAS) recidivism dataset. GitHub Repository URL: https://github.com/propublica/COMPAS-analysis . (2016). 5. D’Angelo, A. FAIR-Edu: Gender Bias in Academic Promotion Dataset. SoBigData Services and Products (2023). https://data.d4science.org/ctlg/ResourceCatalogue/fair-edu_gender_bias_in_academic_promotion_dataset . 6. Wang, Z. et al . Results and insights from diagnostic questions: The neurips 2020 education challenge. In NeurIPS 2020 Competition and Demonstration Track , 191–205 (PMLR, 2021). 7. Choi, Y. et al . Ednet: A large-scale hierarchical dataset in education. In International conference on artificial intelligence in education , 69–73 (Springer, 2020). 8. of Education Sciences, I. National Education Longitudinal Study of 1988. https://nces.ed.gov/surveys/nels88/ (1988). 9. Alshehri, M. et al . Moocrev: A large-scale data repository for course reviews. In International Conference on Artificial Intelligence in Education , 124–131 (Springer, 2024). 10. Le Quy, T., Roy, A., Iosifidis, V., Zhang, W. & Ntoutsi, E. A survey on datasets for fairness-aware machine learning. Wiley Interdisciplinary Reviews: Data Mining and Knowledge Discovery 12 , e1452 (2022). [ Google Scholar ] 11. Fabris, A., Messina, S., Silvello, G. & Susto, G. A. Algorithmic fairness datasets: the story so far. Data Mining and Knowledge Discovery 36 , 2074–2152 (2022). [ Google Scholar ] 12. Meng, L. et al . Less is more: Feature engineering for fairness and performance of machine learning software. ACM Trans. Softw. Eng. Methodol . 10.1145/3730577 (2025). 13. Caton, S., Malisetty, S. & Haas, C. Impact of imputation strategies on fairness in machine learning. Journal of Artificial Intelligence Research 74 , 1011–1035 (2022). [ Google Scholar ] 14. Giovannelli, J. et al . Unfair Inequality in Education: A Benchmark for AI-Fairness Research. 10.5281/zenodo.17592007 (2025). [ DOI ] [ PMC free article ] [ PubMed ] 15. DiSalvo, B., Yip, J., Bonsignore, E. & Carl, D. Participatory design for learning. In Participatory design for learning , 3–6 (Routledge, 2017). 16. Ingels, S. J., Scott, L. A., Taylor, J. R., Owings, J. & Quinn, P. National education longitudinal study of 1988 (nels: 88), base year through second follow-up: Final methodology report. working paper series. (1998). 17. Nakagawa, S. Missing data: mechanisms, methods and messages. Ecological statistics: Contemporary theory and application 81–105 (2015). 18. Little, R. J. & Rubin, D. B. Statistical analysis with missing data (John Wiley & Sons, 2019). 19. Zhang, Y. & Long, Q. Assessing fairness in the presence of missing data. Advances in neural information processing systems 34 , 16007–16019 (2021). [ PMC free article ] [ PubMed ] [ Google Scholar ] 20. Manly, C. A. & Wells, R. S. Reporting the use of multiple imputation for missing data in higher education research. Research in Higher Education 56 , 397–409 (2015). [ Google Scholar ] 21. Harel, O. & Schafer, J. L. Partial and latent ignorability in missing-data problems. Biometrika 96 , 37–50 (2009). [ Google Scholar ] 22. Gelman, A. & Hill, J.Data analysis using regression and hierarchical/multilevel models. New York, NY: Cambridge (2007). 23. Feng, R., Calmon, F. & Wang, H. Adapting fairness interventions to missing values. Advances in Neural Information Processing Systems 36 (2024). 24. Bellman, R. Adaptive Control Processes: A Guided Tour (Princeton University Press, 1961). 25. Sedgwick, P. Pearson’s correlation coefficient. Bmj 345 (2012). 26. Yuan, Y., Wu, L. & Zhang, X. Gini-impurity index analysis. IEEE Transactions on Information Forensics and Security 16 , 3154–3169 (2021). [ Google Scholar ] 27. Mislevy, R. J. Randomization-based inference about latent variables from complex samples. Psychometrika 56 , 177–196 (1991). [ Google Scholar ] 28. Van Buuren, S. & Van Buuren, S. Flexible imputation of missing data , vol. 10 (CRC press Boca Raton, FL, 2012). 29. Rosseel, Y. & Loh, W. W. A structural after measurement approach to structural equation modeling. Psychological methods 29 , 561 (2024). [ DOI ] [ PubMed ] [ Google Scholar ] 30. Rényi, A. On measures of dependence. Acta mathematica hungarica 10 , 441–451 (1959). [ Google Scholar ] 31. Giuliani, L., Misino, E. & Lombardi, M. Generalized disparate impact for configurable fairness solutions in ml. In Proceedings of the 40th International Conference on Machine Learning , ICML’23 (JMLR.org, 2023). 32. Guidotti, R. et al . A survey of methods for explaining black box models. ACM computing surveys (CSUR) 51 , 1–42 (2018). [ Google Scholar ] 33. Ribeiro, M. T., Singh, S. & Guestrin, C. Why should i trust you?: Explaining the predictions of any classifier. In Proceedings of the 22nd ACM SIGKDD international conference on Knowledge discovery and data mining , 1135–1144 (ACM, 2016). 34. Ciatto, G., Sabbatini, F., Agiollo, A., Magnini, M. & Omicini, A. Symbolic knowledge extraction and injection with sub-symbolic predictors: A systematic literature review. ACM Computing Surveys 56 , 161:1–161:35, 10.1145/3645103 (2024). [ Google Scholar ] 35. Pearl, J. Probabilistic reasoning in intelligent systems: networks of plausible inference (Morgan Kaufmann, 1988). 36. Kwisthout, J. Most probable explanations in bayesian networks: Complexity and tractability. International Journal of Approximate Reasoning 52 , 1452–1469, 10.1016/j.ijar.2011.08.003 (2011). Handling Incomplete and Fuzzy Information in Data Analysis and Decision Processes. [ Google Scholar ] 37. Zichao Wang, E. S. P. C. Y. Z. J. M. H.-L. R. E. T. R. G. B. C. B. S. P. J. S. W. C. Z., Angus Lamb. Dataset: Eedi2020 (2024). 38. Feng, W., Tang, J., Liu, T. X., Zhang, S. & Guan, J. Understanding dropouts in moocs. In Proceedings of the 33rd AAAI Conference on Artificial Intelligence (2019). Associated Data This section collects any data citations, data availability statements, or supplementary materials included in this article. Data Availability Statement The data 14 is publicly available on Zenodo https://zenodo.org/records/17592007 ), under the licence Creative Commons Attribution 4.0 International (CC BY 4.0). The structure of the dataset is described in detail in Section; the dataset itself is auto-explanatory, as it contains a README file with all information required to understand how it is organized. The code used to pre-process the collected data is made available at the following webpage: https://github.com/aequitas-aod/Unfair-Inequality-Education . The code is made publicly available under the Apache License, Version 2.0 (January 2004). The main difference with respect to the data repository hosted on Zenodo (see Sec. is that in this latter repository, we also provide the code used to validate the data (the results of the validation are described in Sec.). Articles from Scientific Data are provided here courtesy of Nature Publishing Group ACTIONS View on publisher site PDF (1.2 MB) Cite Collections Permalink PERMALINK Copy RESOURCES Similar articles Cited by other articles Links to NCBI Databases Cite Copy Download .nbib .nbib Format: AMA APA MLA NLM Add to Collections Create a new collection Add to an existing collection Name your collection * Choose a collection Unable to load your collection due to an error Please try again Add Cancel Follow NCBI NCBI on X (formerly known as Twitter) NCBI on Facebook NCBI on LinkedIn NCBI on GitHub NCBI RSS feed Connect with NLM NLM on X (formerly known as Twitter) NLM on Facebook NLM on YouTube National Library of Medicine 8600 Rockville Pike Bethesda, MD 20894 Web Policies FOIA HHS Vulnerability Disclosure Help Accessibility Careers NLM NIH HHS USA.gov Back to Top