Measuring and Mitigating Biases
in Motor Insurance Pricing

The insurance industry is characterized by an inherent reversal in its production cycle, where insurers request a fixed premium at the time of policy subscription in exchange for coverage against uncertain risks in terms of both occurrence and magnitude. This inversion underscores the statistical nature that envelops the pricing of insurance, necessitating adherence to statistical theory for the estimation and coverage of random events. Beyond these statistical considerations, insurance premiums represent the equitable price for insurance services, encompassing a multitude of strategic challenges. The insurance market is increasingly competitive, comprising established incumbents and more agile newcomers striving to gain market share. This competitive landscape is further catalyzed by continuously evolving regulations aimed at fostering competition among industry players.

Hence, participants in the insurance market must offer competitive pricing strategies that align with their corporate strategies and communication, while also employing tools tailored to diverse distribution processes.

Recent years have witnessed the widespread adoption of machine learning algorithms, neural networks, and the utilization of vast datasets. This adoption is attributable to scientific advancements, increased computational power, enhanced accessibility to technology, and the proliferation of data. These emerging technologies have made significant inroads into various business sectors, including the insurance industry.

These data and algorithms serve as decision support tools, aiding in policyholder segmentation, risk comprehension, and the consideration of various factors associated with it. Therefore, industry stakeholders must incorporate these new elements to maintain their competitiveness. Nevertheless, the extensive use of massive data and intricate algorithms has brought the issue of transparency to the forefront. It is imperative that premiums and decisions are both explainable and fair, given the high stakes involved. This is not simply about algorithms assisting in trivial choices like movie selection; rather, it involves algorithms determining the cost of access to insurance services for various population segments. Concerns regarding fairness and ethics have been integral to our societies and philosophies for centuries. Although subject to interpretation, fairness can be defined as the ability to place individuals on an equal footing while acknowledging the differences that exist among them.

In response to these considerations, regulatory measures such as the gender directive have been implemented to promote fairness. Reforms related to access to borrower insurance also represent a form of fairness enforcement. Consequently, insurance industry participants may need to construct fairer premiums with regard to so-called sensitive or protected variables. These fairness constraints can emanate from regulatory requirements, as seen in the case of gender, or be driven by commercial and strategic objectives, such as those related to age in specific companies.

In the context of this study, bias is understood as a form of discrimination, signifying the undesirable impact of a sensitive variable on a variable of interest. For instance, this could include the effect of gender on insurance premiums within the framework established by the gender directive (of the European Union (2008)). Since 2016, numerous research studies (Angwin et al. (2016); Chouldechova (2017))have identified instances of discrimination in decision-making tools across various domains, such as the risk assessment tool for recidivism in the United States, Google Images algorithms, Amazon’s application processing algorithms, and lending and financing algorithms, among others. Historically, the solution has often been to circumvent the issue rather than address it from an ethical perspective.

One simplified scenario that can be examined involves the use of a gender-correlated variable to determine premiums. In this case, gender is not explicitly factored into the estimates, resulting in fairness by omission. However, the presence of a correlation between gender and the non-sensitive explanatory variable leads to an indirect effect of gender on the estimated premiums. Even though gender is seemingly reprocessed or removed from the model, its association with other variables allows its influence on the target variable to persist. This is because adjustments only account for the direct impact of sensitive variables and not their indirect effects, despite these variables exerting a significant influence on the distributions of other explanatory variables. Factors such as age, gender, and disability influence choices related to activities, risk tolerance, product preferences, and so on. Consequently, it is imperative to first establish the means to define and measure the fairness of the constructed models and subsequently mitigate this bias through fairness-aware approaches.

The purpose of this article is to provide actuaries with tools to understand, measure and mitigate the unwanted effect of a sensitive variable in a pricing problem.In Section 2, we will discuss fairness notions and statistical measures of fairness. The bigger picture will be shown while focusing on important elements for pricing. Then, in Section 3, we will present methods that can be used to reduce fairness bias in pricing models. Finally, in Sections 4 and 5, we present the results of fairness implementation on a real car insurance pricing, respectively with measures of biases (in Section 4) and a description of the impact of mitigation (in Section 5). Note that all implementations are made in Python 3.8.

Throughout this document, the variable denoted as $Y$ is the target variable we aim to predict. This variable can take on either categorical or quantitative values (discrete or continuous, but positive). Given the primary focus on pricing, $Y$ predominantly assumes a quantitative nature. However, for the sake of illustration, it might be temporarily treated as a binary variable. In the context of addressing this supervised machine learning problem, we introduce an algorithm denoted as $m$, a set of non-sensitive features represented by the vector $\boldsymbol{X}=(X_{1},X_{2},\dots,X_{p})$, and sensitive attributes denoted as $\boldsymbol{S}$. The vector $\boldsymbol{S}$ comprises variables that are intended to have no influence on the models, either intentionally or inadvertently. These variables are typically discrete, particularly in the context of examining fairness in pricing, and may explicitly pertain to attributes such as gender. If a single sensitive attribute is considered, as it will mostly be the case in this document, notation $S$ will be used (see Hu et al. (2023) for a detailed discussion about multiple sensitive attributes). Furthermore, we assume that $S$ is a binary sensitive variable, with possible values of $0$ and $1$, which may correspond to the gender of the policyholder.

Each data point in our dataset, which can represent individuals, contracts, or claims, is identified as $(y_{i},\boldsymbol{x}_{i},s_{i})$, with $i$ ranging from $1$ to $n$, and $n_{0}$ and $n_{1}$ signifying the numbers of observations where $s_{i}=0$ and $s_{i}=1,$ respectively. Note that lowercase letter variables denote observations, whereas uppercase letter variables denote random variables.

We introduce $\widehat{y}=\widehat{m}(\boldsymbol{x})$ as the predictions generated by an "unfair model" (or "unaware model", as defined by Dwork et al. (2011, 2012)), and $\widetilde{y}=\widetilde{m}(\boldsymbol{x})$ as the predictions generated by a fair model. $\widetilde{m}$ is a model that takes into account defined fairness constraints to minimize the effects of $s$ on $y$. As studied in section 3, these constraints can take the form of criteria for variable selection (preprocessing constraints), penalized learning (in-processing approach) and output correction (post-processing approach). We can then define individual fairness bias, denoted as $\varepsilon$, as the difference between these two predictions, i.e., $\varepsilon=\widehat{y}-\widetilde{y}$.

We also introduce the function $\mathcal{V}^{s}(\boldsymbol{x}_{i})$, which represents the $k$-nearest neighbors in the group $S=s$ associated with the $i$-th individual.

Group fairness necessitates equality of treatment among groups based on the sensitive variables. For that, statistical properties are specified using the model $m$ and $\widehat{Y}$ to compare each group. The overarching objective is to ensure that individuals from both privileged and unprivileged groups are treated equitably in terms of the specified statistical properties.

Fairness was initially introduced using conditional probabilities in the context of binary classification (see EEOC (1979)).This concept can be extended to scenarios where the variable $Y$ is non-binary, incorporating moment properties, a weak version, or distribution properties, a strong version. These can be linked to correlation properties and independence conditions, respectively.

Individual fairness, also known as similarity-based fairness, asserts that similar individuals should receive similar predictions. Unlike group fairness, which aggregates outcomes at the group level, individual fairness compares individuals directly. This concept is formalized as "disparate treatment":

Individual fairness and disparate treatment embody an intuitive conception of fairness, seen in scenarios like car insurance pricing, where identical individuals with different sensitive attribute values receive the same predictions. However, this approach overlooks relationships between explanatory variables ($\boldsymbol{X}$) and $S$, impacting fairness enforcement.

The challenge of individual fairness arises from sensitive variables significantly affecting other variables’ distribution. Attributes like gender, age, disabilities, ethnicity, and illnesses shape habits, preferences, and behaviors, complicating defining "similar individuals."

To address these questions, Dwork et al. (2011) introduced the Lipschitz constraint concept:

Here, $d_{Y}$ and $d_{\boldsymbol{X}}$ measure distances in target and explanatory variable spaces, respectively. Similarity in the $Y$ space, like between premiums, can be defined using metrics such as $|\widehat{y}_{i}-\widehat{y}_{j}|$ or $(\widehat{y}_{i}-\widehat{y}_{j})^{2}$. The focus is on defining individuals as similar when they have different values of the sensitive attribute ($S$). And, disparate treatment will mean that regardless of the value of $S$, if two individuals are similar on $\boldsymbol{X}$, they must be similar on $Y$ and thus have the same premium.

Despite its intuitive appeal, individual fairness is less frequently used than group fairness, largely due to addressing various interactions between variables’ complexity. Causality-based criteria may offer a promising avenue, but practical implementation remains challenging. These causal models are not discussed here (for more information, refer to Kusner et al. (2017), Alycia and Wu (2022), and Galles and Pearl (1998)).

In light of the definitions provided earlier, it is imperative that fairness metrics possess the ability to quantify full/conditioned independence in the context of group fairness and also discern discrepancies in individual predictions, contingent upon the proximity of individuals in the case of individual fairness.

These mitigation methods involve data transformations aimed at reducing bias while retaining relevant information. The methods we have implemented include: total removal of variables correlated with the sensitive variable (Section 3.1.1), removal of linear correlations (Section 3.1.2) and an adaptation of the fair-SMOTE method (Section 3.1.3) proposed by Chakraborty et al. (2021). Originally designed for binary cases, we have modified it to suit the needs of bias mitigation in non-life insurance pricing.

These approaches involve the inclusion of fairness constraints during the model calibration phase. The exponentiated gradient method, named after the game theory technique upon which it is founded, and its grid search version are both utilized.

This approach involves transforming model predictions to enhance fairness. For instance, logistic regression allows influencing the model’s behavior with respect to different classes through computed probabilities and decision thresholds without recalibrating predictions.

Historically, post-modeling techniques mainly adjust decision boundaries while considering fairness definitions, but they’re less applicable to the continuous case due to the absence of distinct decision boundaries. Recent suggestions, like using Wasserstein barycenter on scores and the associated transport procedure by Charpentier et al. (2023b), aim to address this gap. However, defining advantageous and disadvantageous premiums is challenging, as it depends on policyholder characteristics and inherent risk. The "individual fair redistribution" approach was introduced within this context.

Table 3 displays the results of the dependence measures between each of the variables of interest and the sensitive variable.

Kendall’s tau indicates a weak dependence between the variables of interest ($Y$ and $S$), with a negative sign suggesting slightly weaker values for women. HGR also detects slightly stronger relationships, consistent with Kendall’s tau. The KS test highlights significant differences between male and female distributions.

According to the flip-test, compared to similar male policyholders, women have an average cost €10 lower. This observation aligns with the negative sign of Kendall’s tau and the mean ratio values, indicating lower averages for women compared to men.

Figures 1 and 2 analyze the distributions of these variables with respect to gender. For the frequency variable, distributions between $[0,1)$ for males and females are similar, representing 97.3% of the population. At $1$, there’s a 15% higher representation of women compared to men, accounting for 1.2% of the population. Above $1$, men are more strongly represented, with an 18% higher presence compared to women, representing 1.5% of the population.

The observed distributions suggest that, on average, men generally exhibit higher values for claim cost, frequency, and premium compared to women in historical data. However, interpreting these figures in absolute terms poses challenges due to the difficulty in establishing critical thresholds, given the lack of precedents in this type of problem. Nonetheless, these initial findings imply that gender has an impact on the variables of interest. Besides the class imbalance (more men than women), bias appears evident in the historical data. Such differences are well-documented in automobile pricing and are typically addressed by either excluding gender as a factor or rebalancing the model outputs. It’s worth noting that risk exposure may differ between gender classes due to various factors (see Ayuso et al. (2016)). To mitigate this effect, our dataset comprises individuals who subscribed to the same mileage package, providing the most accurate information available on vehicle usage.

By definition, the best-performing model would include gender as a feature because it leverages all the available information for risk modeling. However, it is also the most unfair model concerning gender since a clear distinction between men and women is directly visible in the resulting premiums. To create this model, gender is reintroduced into the modeling process. Once constructed, predictions $\widehat{Y}$ are obtained, and the dependence between $S$ and $\widehat{Y}$ is measured. Table 4 presents the results of these different measures.

The dependencies between the interest variables and the sensitive variable were amplified according to almost all measures. Measures leveraging distributions, like div_JS and the flip-test, suggest weaker dependencies because $\widehat{Y}$ has less dispersion than $Y$. Kendall’s tau detects triple the dependence for the cost variable, implying a somewhat simpler detection of the dependency structure. Thus, the constructed models not only replicated the historical bias present in the data but also exacerbated it. Figures 3, 4, and 5 analyze the distributions of these predicted interest variables concerning gender.

The mismatch between the distributions is more pronounced post-modeling, with larger portions of the distributions not overlapping. The differences are more evident compared to the historical data case.

How can we explain this bias amplification? The interpretation of the constructed models reveals that gender plays a significant role in the prediction processes. By examining coefficients and importance measures, it becomes evident that gender is among the most important variables in these models.

The greater measured bias in the constructed models compared to the historical data used for modeling can be explained by the interdependence of the variables in the dataset. There is a link between the explanatory variables selected and the variables of interest because they help understand the associated risk. However, these variables, besides their predictive abilities, may also be related to each other. For example, a vehicle’s power may be linked to its price. These interdependencies pose challenges for fairness implementation because the so-called sensitive variables often have a notable influence on the distributions of the other observed variables. Thus, even if the observable interdependencies seem weak at first glance, their accumulation can magnify the role of the sensitive variable in the models.

In the data used, gender is significantly linked with variables such as horsepower, weight, gearbox type, and vehicle price, in addition to its minor relationships with other explanatory variables. Figure 6 visually represents how gender directly influences the variables of interest and has a significant indirect effect through its relationships with other variables.

For instance, in the cost model, gender is ranked as the fourth most influential variable. Nevertheless, it’s noteworthy that the most substantial variables within the model are those closely related to gender. Generally, sensitive variables exert a significant impact on individuals’ behavior. For instance, age can influence risk-taking propensity, the level of responsibility, personal interests, and more. Similarly, ethnicity may correlate with factors such as geographical location, purchased services, and cultural preferences.

These analyses inevitably lead to a central question: What are the implications when gender is omitted or excluded from the modeling process? This raises the possibility that the observed interdependencies might have been overemphasized, and that the models have unduly relied on gender as a driving factor.

By definition, a model that excludes gender as a variable aims to eliminate direct discrimination, ensuring that final outcomes are not discerned based on gender. However, it’s acknowledged that this strategy doesn’t address indirect discrimination as it overlooks the interrelationships between other explanatory variables and gender. Table 5 presents findings on the association between $\widehat{Y}$ and gender in models where gender isn’t included as an explanatory variable.

The results show a minimal reduction in bias compared to models where gender is included, indicating that gender’s influence remains pronounced even when it’s absent from the models. While it may seem counterintuitive that premiums no longer differentiate by gender, there are still distinctions observed based on gender, primarily due to indirect interdependencies within the data. For example, varying premiums based on the presence of secondary drivers indirectly leads to varied premiums for each gender because secondary drivers have different distributions by gender. In complex multivariate models, such distinctions based on gender can arise from the combination of multiple variables. For instance, a combination of factors such as vehicle power, gearbox type, and zoning may result in significant gender imbalances within the dataset. When examining the intersection of zoning and vehicle prices, notable gender imbalances become evident in the data. For instance, 82% of individuals residing in the 10th zone and owning a vehicle costing over €30,000 are male. Consequently, any substantial premium difference within this segment compared to others significantly impacts men, leading to gender-based premium disparities, despite the absence of explicit gender classification.

These findings highlight that implementing fairness through omission does not offer a comprehensive solution to discrimination. An alternative modeling approach is employed to assess dependency levels. Instead of excluding the gender variable, it is included in the model, and then the outputs are post-processed to ensure gender-insensitive pricing, aligning with strategies used by insurers to comply with regulations like the gender directive. The specific approach calculates a weighted average of male and female model outputs for all segments, but the results show similar dependencies between the new values of $\widehat{Y}$ and $S$ as previous modeling strategies. This method fails to resolve interdependencies, as reweighting the outputs perpetuates existing imbalances.

Consequently, gender’s influence persists within the modeling process, even after its omission or conventional adjustment. Some scholars refer to this as the "reconstruction" of the sensitive variable, where bias continues to manifest due to interdependencies between explanatory variables and the sensitive variable. This study exemplifies such a scenario.

In summary, bias exists in historical data, and regardless of the modeling approach applied, it is perpetuated, primarily due to interdependencies. The subsequent section explores bias mitigation while maintaining acceptable performance, using the model constructed without gender as reference.

To establish the deletion scenarios, we analyze the relationship between sensitive and non-sensitive variables. The results are depicted in Figure 7.

By examining these dependency levels, $\alpha$ can assume values within the range of $(8\%,41\%].$ Scenarios are developed by iteratively adjusting the threshold, considering the relationships between the variables. The chosen scenarios are as follows:

• •

  Scenario 1: driv_2;

• •

  Scenario 2: driv_2 and box_type;

• •

  Scenario 3: driv_2, box_type and energy;

• •

  Scenario 4: box_type, driv_2, energy and weight;

• •

  Scenario 5: box_type, driv_2, energy and anc_cp;

• •

  Scenario 6: box_type, driv_2, energy, and zoning;

• •

  Scenario 7: box_type, driv_2, energy and veh_price;

• •

  Scenario 8: box_type, driv_2, energy, veh_price and weight_kw;

• •

  Scenario 9: box_type, driv_2, energy, veh_price and zoning;

• •

  Scenario 10: box_type, driv_2, energy, weight and zoning;

• •

  Scenario 11: box_type, driv_2, energy, veh_price, weight_kw and zoning.

The gender variable is eliminated from all models. Figure 8 illustrates the fairness and performance levels of each scenario. The non-dominated scenarios are highlighted in red. The abbreviation dbe refers to the variables driv_2, box_type and energy. For example, scenario 6 is represented on the graph as dbe+zoning.

Among the non-dominated scenarios, a choice must be made based on the desired trade-off as determined by decision-makers. For instance, the objective might be to attain the fairest model while tolerating a maximum performance loss of 5%. Scenarios 8, 9, 10, and 11, although among the fairest, would no longer meet this criterion. Upon analyzing the graph, the 7th scenario, labeled as dbe+veh_price, emerges as an interesting trade-off option, as it significantly reduces bias, almost halving it, with an acceptable performance loss. Compared to the reference model, a bias reduction of 37% is achieved in exchange for a 3.9% drop in performance. The 7th scenario model is further examined.

It appears that the model has learned to rely more on other variables to compensate for the absence of the veh_price variable. For instance, some variables like zoning, driv_ly, and veh_age now have larger coefficients and play a more significant role in the prediction process. However, the performance loss still indicates that these omitted variables cannot be entirely replaced optimally. Furthermore, this new model leads to a slightly degraded equilibrium in pricing compared to the reference equilibrium. The LR metric obtained is 99.3% compared to the reference LR of 99.7%. It is essential to note that once the input data are modified, a reevaluation and validation of the model are necessary.

Hence, this initial approach can yield satisfactory results depending on the quality of the explanatory variables and the acceptable level of performance loss. It is a straightforward method to implement and underscores the importance of addressing fairness concerns from the initial stages of data processing and variable selection.

To apply the methodology introduced in Section 3.1.2, we have chosen to retain qualitative variables while focusing solely on transforming quantitative variables. The quantitative variables subject to transformation include: veh_price, weight_kw, veh_age, driv_yp, and driv_ly. We uniformly select $100$ values of $\alpha$ within the interval $[0,1]$, including both endpoints. For each $\alpha$ value, we correct our quantitative variables, build our model, and measure performance and fairness.

From these results, two observations arise. First, there is no clear trend between the fairness level and the hyperparameter $\alpha$. Typically, one might expect fairness and performance to decrease as $\alpha$ increases, but this trend is not evident. The relationship between fairness, performance levels, and $\alpha$ appears random, lacking a discernible pattern. Second, using this method, some models outperform the reference model. For example, the model with $\alpha=85\%$ shows lower bias and slightly better performance. While the performance gain is only around 0.5%, it comes with an 11% improvement in fairness. However, the efficacy of this method relies on the linear model’s ability to detect relationships, ensuring that residuals are unbiased. Nonetheless, examining linear correlations reveals that the highest dependence between $S$ and other explanatory variables is 8%, significantly limiting suppression capacity in this specific case.

While this method seems promising, it presents several limitations within the insurance pricing context. First, in pricing models, it is more common to use discrete or discretized variables as inputs. Ideally, continuous variables should be discretized after suppression to ensure interpretability. However, as the values represent residuals, they may lose interpretability. Additionally, in practical production scenarios, transforming customer attributes into predicted premiums may be complicated by this method. Lastly, the non-intuitive model behavior in response to changes in $\alpha$ lacks a clear explanation. Considering these factors, while a model outperforming the reference model was achieved, adopting it is challenging due to associated limitations. This method may find greater success when applied to a continuous sensitive variable.

Balancing the number of bins and the size of each subpopulation, we have opted for a segmentation into seven bins, with the following breakdown:

• •

  bin 1 if $y=0$;

• •

  bin 2 if $y\in(0,250]$;

• •

  bin 3 if $y\in(250,500]$;

• •

  bin 4 if $y\in(500,750]$;

• •

  bin 5 if $y\in(750,1000]$;

• •

  bin 6 if $y\in(1000,1500]$;

• •

  bin 7 if $y\in(1500,+\infty)$.

The number of bins serves as a hyperparameter, resulting in various binning scenarios. The selection of these bins depends on their alignment with the observed trends in the premium distribution.

Following the bin selection, the next step is to determine the scope of resampling. As previously discussed, resampling will only be performed on $S|Y$ since resampling on $Y$ would alter the specific distribution needed for the GLM. Although for algorithms like random forest, which make no distribution assumptions, resampling on $Y$ was attempted, it led to a notably lower LR due to the high premium levels predicted by the models.

It’s worth noting that the creators of the traditional fair-SMOTE method recommend setting $st$ and $ft$ to 0.8 each. This configuration is retained for consistency. The distribution after resampling is illustrated in Figure 9.

In total, 9588 individuals were generated, constituting a substantial expansion of the learning dataset by 16.1%. The resulting distribution of this enlarged dataset remains in alignment with the initial observations, and the distinctions between genders are notably less prominent. The box plots demonstrate a substantial overlap, with the noticeable shift observed in the historical data significantly mitigated. Therefore, this resampling procedure appears to be effective. Following the preparation of the augmented training data, the models are reconstructed, and the outcomes are presented in Table 6.

The initial distribution shows reduced skewness, but the outcomes on the test dataset still exhibit skewed results. The improvement in fairness is minimal, with less than a 4% enhancement. Performance metrics and the LR remain similar to the baseline. Therefore, rebalancing the distributions solely based on $S|Y$ seems inadequate for achieving fairness. This could be because the bias needing mitigation in car pricing does not stem from a representation issue concerning $Y$, but rather from historical differences in male and female premiums. Thus, while resampling ensures both genders receive sufficient data for equitable treatment during training, it doesn’t address the interdependencies between variables, crucial for ensuring fairness.

Sensitivity analyses on various modeling choices were conducted to explore whether these choices contributed to the ineffectiveness of our approach. Changes to hyperparameters $ft$ and $st$ didn’t result in improved outcomes. Adjusting the resampling bins also didn’t lead to significant enhancements. Increasing the number of bins introduced more irregularities in the distribution of $Y$ and reduced model performance.

A significant departure from conventional practices in the literature was limiting the resampling process solely to $S|Y$ rather than conducting full resampling on $Y$. Full resampling of $Y$ altered its distribution drastically, generating an excessive number of artificial instances and failing to produce satisfactory results. To explore the potential benefits of resampling on $Y$ while preserving the overall distribution shape, partial rebalancing of $Y$ values was considered. For instance, in scenarios where $Y$ has only two distinct values, such as $\{25,35\}$, Tables 10 to 10 demonstrate the various forms of resampling undertaken.

Multiple scenarios involving partial resampling of $Y$ were explored. The results indicate that as the number of artificially generated individuals increases, model performance worsens without simultaneous improvements in fairness. For example, expanding the population by 20% through resampling led to an increased RMSE of 177, while the HGR fairness metric remained at 28.05. Ultimately, these approaches failed to produce better results.

An unexplored area for potential improvement involves considering the explanatory variables during the resampling process. Instead of resampling solely to achieve equity based on $S|Y$, an alternative approach is resampling based on $S|X_{1},\dots,X_{p},Y$. The selection of variables for conditioning is crucial here, as they can potentially amplify existing biases. Implementing this approach is complex, as it requires overseeing the quality of resampling on individual variables and their interactions while maintaining the overall consistency of premium distributions. Although this approach hasn’t been extensively explored in this study, integrating more efficient methods for addressing interdependencies between variables may be considered in future research.

In summary, the various pre-modeling bias mitigation strategies have provided insights into the data’s structure and identified elements with significant influence on bias levels. They offer valuable tools for addressing ethical concerns during the data preprocessing phase.

To achieve the model with the lowest permissible error level, bounded by $\zeta$, a specific approach is adopted. The method starts with the lowest feasible $\zeta$ value, which is initially set at $10^{-5}$, and progressively increases it in multiples of 10 until the algorithm converges. The first convergence occurred when $\zeta=10^{-1}$. Once the method converges to the optimal solution, the grid search and the $M$ parameter become unnecessary.

However, for $\zeta<10^{-1}$, these tools can help identify suboptimal solutions that may outperform the solution found with larger tolerances. During testing, values of $M$ ranging from 0.5 to 500 were employed with grid sizes of 3000. The results obtained were not as favorable as those achieved with convergence. It is essential to note that the grids explored were relatively small compared to the dimensionality of the dataset. Given more computational power and time, improved results might potentially be attainable.

The outcomes for the best model are presented in Table 11.

The model obtained after mitigation exhibits reduced efficiency and increased unfairness compared to the reference model. While the loss of performance is negligible, the model experiences a 9% higher HGR after the application of the mitigation method. These outcomes can be attributed to the fact that the error constraint implemented does not enforce any form of independence between $\widehat{Y}$ and $S$. In other words, the error rate per gender can be the same without ensuring fair treatment of the genders.