In the last decade, data-driven decision-making is considered one of the main drivers for organizational success. Within this approach, decisions are based on insights and patterns identified through data analysis. In this scenario, input data must be reliable to guarantee the accuracy of the results: they should be correct and complete but also unbiased, i.e., both Data Quality (DQ) and Fairness should be guaranteed. However, maximizing DQ and Fairness simultaneously is not trivial, since data quality improvement techniques can negatively afect Fairness and vice versa. Understanding and thoroughly analyzing this relationship between DQ and Fairness is therefore paramount, and is this paper's goal. The results of our experiments, based on a well-known biased dataset (the Adult Census Income) provided details about this trade-of and allowed us to draw some guidelines.
In the last decades, the possibility to use large amounts of data to extract information, and
gain deeper knowledge in several domains, has caused the spread of a data-driven culture,
making data collection and management extremely important. In fact, in this scenario, strategic
decisions are made on the basis of data analysis and interpretation, and relying on dependable
results becomes imperative. The performance of Machine Learning (ML) algorithms may be, for
example, seriously afected by the poor quality of the training data [
Another key property for efective data-driven decision-making applications is the ethical
level of the data. In fact, even the most accurate application for collecting data might be afected
by ethical issues, since also high-quality data might lead to unfair outcomes. In [
Looking at this list, it is immediate to see that there may be contrasting objectives also among the
dimensions of Ethics, for instance as regards an obvious conflict between Transparency and Data
Protection. In the same way, the relationship between many well-known DQ dimensions [
With this work, we want to investigate this trade-of, especially focusing on the Completeness and Accuracy dimensions of DQ, and on the Fairness dimension of ethics. To this aim, we have designed experiments that take as input a dataset and perform an assessment of DQ and Fairness before and after the application of some operations that should improve them. After the description of the experiments we conclude the paper with some takeaway messages for the researchers that look for the best strategy for applying changes in data to improve Fairness or DQ according to their needs.
The rest of the paper is organized as follows: Section 2 summarizes related work, while Section 3 introduces preliminary concepts of both areas of DQ and Fairness and describes the method we used to analyze the relationship between DQ and Fairness; Section 4 presents the experiments we conducted on a real-world dataset, and Section 5 concludes the paper.
Research studies on the relationship between DQ and Fairness are in a very preliminary phase. In this section we will first present seminal works in Fairness and then introduce two first attempts at studying this important relationship. We do not focus on DQ systems since in this paper we will resort to well-known and established DQ techniques.
In the literature, one of the most notable solutions is AI Fairness 360 [
Similarly, Fairlearn [
A preliminary system that considers both DQ and Fairness is the paper by Abraham et al. [
A similar system has been presented by Biswas et al. [
This section presents the method we used to investigate the relationship between DQ and Fairness. Before describing the work, we introduce some preliminary theoretical concepts related to various Data Quality and Data Ethics aspects.
Data Quality Data Quality (DQ) is defined as “fitness for use,” i.e., the ability of a data
collection to meet the user requirements [
Fairness Fairness is one of the most important dimensions of Data Ethics. The most used definition of Fairness is: “it is the absence of any prejudice or favoritism toward an individual or a group based on their inherent or acquired characteristics” [10, p.100].
Fairness is based on the idea of protected or sensitive attribute. A protected attribute is a
characteristic for which non-discrimination should be established, such as religion, race, sex,
and so on [
There is no unique definition of Fairness, but many facets exist, and a model is considered
fair if it satisfies some or all these definitions. The most used technique to identify unfairness
in datasets is to train a classification algorithm to predict the binary value of the target class
and then use Fairness metrics to understand whether the prediction of this model encompasses
discrimination: if the metrics results show unfairness, we can conclude that also the original
dataset contains unfair behaviors, since the model learned the bias from it. Specifically, we
measure the importance of protected attributes in determining the result of the model. The
following statistical metrics study how specific values of the protected attributes impact the
result of the prediction algorithm (e.g., being a woman determines that the salary is lower than
50k$/year, and being a man determines that the salary is higher than 50k$/year). Informally:
• Disparate Impact is the probability to get a positive outcome regardless of whether the
person is in the protected group [
The Method Adopted This section presents the two pipelines we defined to execute the experiments and study the relationship between DQ and Fairness. In the first one, we inject errors in the dataset, causing data quality issues, then apply DQ improvement techniques and measure their impact on Fairness. In the second pipeline, since the Adult Census Income dataset1 already contains bias w.r.t. the income of US citizens we do not need to inject further bias to perform the experiments, therefore, we already start from a biased dataset, apply bias mitigation techniques and measure their impact on DQ. By means of these results, we evaluate the relationship between Fairness and DQ.
The dataset used for the experiments had the following characteristics: (i) it contained bias, so the classification algorithm would be afected by unfairness, (ii) it had been pre-processed so that the classification algorithm could be executed in a correct manner. For example, missing values had to be dealt with, and encoding and normalization operations were performed.
The last operation to be performed before entering the two pipelines was applying a first
classification algorithm in order to compute the Fairness level of the dataset. As for the DQ
measure, we already knew that it was 100%. We now describe the two pipelines.
• Data Quality Improvement Pipeline: The input dataset is free of DQ problems. For this
reason, we had to inject errors in order to evaluate the impact of DQ improvement
1https://archive.ics.uci.edu/ml/datasets/adult
techniques. By injecting in a uniform manner a diferent percentage of DQ errors (from
90% to 0%, with a decreasing step of 10%) related to a specific DQ dimension, the Error
Injection phase generated ten instances of the original dataset at diferent levels of quality.
For example, for the Completeness dimension, a certain number of values are replaced
with null values. The obtained ten dirty datasets were the input of the Data Quality
Improvement phase, in which a DQ improvement technique was applied. In the case of the
Completeness dimension, an imputation technique was selected. The ten clean datasets
obtained as output were analyzed in the second Evaluation phase, in order to check the
impact of the DQ improvement on the Fairness measures. This procedure was repeated
for a number of diferent imputation methods.
• Fairness Improvement Pipeline: As regards Fairness, we did not have an error injection
phase since the considered database was already biased. The improvement phase (i.e., Bias
Mitigation phase) consisted of applying a bias mitigation technique to remove unfairness.
The repaired dataset, output of this phase, was analyzed in the second Evaluation phase
in order to assess the impact of the bias mitigation on the DQ level. This phase was
repeated for all the selected bias mitigation techniques. Since some of these techniques
act by directly substituting the data values with other (faked) values, they also allow
for partial bias repairs. For example, Correlation Remover [
The last step was the analysis of the results of the two second Evaluation phases.
In this section, we first introduce the experimental setup and then describe the results of the experiments, both from the DQ and the Fairness perspectives.
DQ Improvement phase In this paper, we consider two types of Data Imputation techniques:
Density-based, i.e., missing values are imputed for each feature with the same distribution of
the non-empty values, and Rare-based, i.e., the less frequent value is imputed.
Bias Mitigation phase Two bias mitigation techniques are proposed in order to remove
the unfairness from data. The former one, Correlation Remover [
Accuracy = ℎ where is the total number of cells.
The three metrics selected to evaluate Fairness (see Section 3) were applied by using the following formulas: (1) (2) (3) (4)
Disparate Impact (DIR) =
Predictive Parity (PPR) = False Positive Ratio (FPR) = (ˆ = 1| = ) (ˆ = 1| = ) ( = 0|ˆ = 1, = ) ( = 0|ˆ = 1, = ) (ˆ = 1| = 0, = ) (ˆ = 1| = 0, = ) where: • : protected attribute that has two values discr (=discriminated), priv (=privileged); • : additional data regarding the individual; • : actual classification result, two values (or labels) 0 or 1; • ˆ : the algorithm predicted decision for the individual, two values of the outcome 0 or 1. Dataset and classification algorithm The Adult Census Income dataset is obtained as an extraction of the 1994 U.S. Census database. It is typically used to predict whether the income of an individual exceeds 50k$ per year. It comprises 48842 tuples, described by 15 attributes, including the targeted class. There are some sensitive attributes, such as ‘race’, ‘sex’, and ‘native country’. In particular, for the experiments shown in this paper, we compute the Evaluation considering the sex as a protected attribute (see Section 3). Finally, the Decision Tree Classifier ofered by the scikit-learn2 Python library was used as classification algorithm .
This section presents the main results we obtained. In Figure 1, the x-axis represents the Completeness level; instead, in Figure 2, the x-axis shows the degree of bias mitigation. In both Figures, the y-axis represents the level of the evaluated metrics. Data Quality Perspective The plots shown in Figure 1 focus on the DQ perspective in which the Evaluation results are compared for the two imputation techniques explained above (see Section 4.1). As we can notice, the Density-based Imputation is widely better: the Accuracy measure does not reach less than 50% for the Accuracy assessment, and the Fairness measures increase as the percentage of injected errors increases. This is related to a vast majority of the target class that has a value lower than 50k$/year in the dataset; since the imputation follows the value distribution, it means that those labels have a higher probability of being assigned to men (who are over-represented). In this way, the dataset will be balanced. We can conclude that the application of this Imputation method improves Fairness. Instead, applying the Rare-based Imputation, we have a big deterioration of Accuracy, and from 100% to 40% of Completeness, the Fairness increases; instead, for Completeness values below 40%, Fairness decreases very quickly. In this specific case, this happens because by imputing the less frequent values, the dataset will be more balanced in favor of the protected classes. As the percentage of injected errors increases, the rare values become too many, unbalancing the dataset. We can also observe that the Predictive Parity (PPR) metric can assume values greater than 1. This means that the privileged class (men) is discriminated for that specific Fairness aspect; False Positive Ratio (FPR) always takes opposite values with respect to PPR. The two metrics are symmetrical since they represent opposite Fairness aspects. From these results, we notice that a trade-of between DQ and Fairness is present and that, from the DQ perspective, this trade-of can be more or less emphasized depending on the DQ improvement technique applied.
Fairness Perspective The plots shown in Figure 2 focus on the Fairness perspective. The
Evaluation results are compared for two diferent bias mitigation techniques: Correlation
Remover ofered by Fairlearn [
Takeaway message From our experiments, we can notice that in some particular cases, the bias mitigation technique that less afects the DQ is not the one that improves Fairness the most, and vice-versa; for these cases, we can deduce that techniques that succeed in preserving both DQ and Fairness do not exist. Therefore, as a takeaway message, we can afirm that the best DQ improvement/bias mitigation techniques to apply strictly depends on the analysis goal. If a user is more interested in preserving Fairness aspects, s/he will concentrate on a subset of techniques at the cost of losing DQ; if the major interest is to optimize the improvement of the DQ, the user will apply a subset of DQ improvement tasks that could deteriorate Fairness. It is worth noting that situations may also exist in which the DQ and the Fairness aspects are not in conflict; however, this is strictly context-dependent and could be rare to observe.
In this work, we analyzed the relationship between Data Quality (DQ) and Fairness. In particular,
through a series of experiments, we demonstrated that between DQ and Fairness, a trade-of
is present. In fact, the experiments showed us that the application of Fairness improvement
operations could lead to a deterioration of the DQ and vice-versa. Analyzing the experiments
more in detail, we can also state that the amount of DQ deterioration after Fairness
improvements depends on the bias mitigation technique, as well as the deterioration of Fairness can
depend on the selected DQ improvement technique. Future work will focus on the definition
of clear guidelines to recommend the best choice of DQ improvement/bias mitigation
techniques to be applied depending on the scope of the analysis. Moreover, we could enrich the
gathered knowledge with more datasets, DQ dimensions, Fairness metrics and bias mitigation
techniques [