Fairness in Predictive Marketing: Auditing and Mitigating Demographic Bias in Machine Learning for Customer Targeting

Abstract

As organizations increasingly turn to machine learning for customer segmentation and targeted marketing, concerns about fairness and algorithmic bias have become more urgent. This study presents a comprehensive fairness audit and mitigation framework for predictive marketing models using the Bank Marketing dataset. We train logistic regression and random forest classifiers to predict customer subscription behavior and evaluate their performance across key demographic groups, including age, education, and job type. Using model explainability techniques such as SHAP and fairness metrics including disparate impact and true positive rate parity, we uncover notable disparities in model behavior that could result in discriminatory targeting. We implement three mitigation strategies—reweighing, threshold adjustment, and feature exclusion—and assess their effectiveness in improving fairness while preserving business-relevant performance metrics. Among these, reweighing produced the most balanced outcome, raising the Disparate Impact Ratio for older individuals from 0.65 to 0.82 and reducing the true positive rate parity gap by over 40%, with only a modest decline in precision (from 0.78 to 0.76). We propose a replicable workflow for embedding fairness auditing into enterprise BI systems and highlight the strategic importance of ethical AI practices in building accountable and inclusive marketing technologies. technologies.

1. Introduction

In the digital economy, machine learning (ML) models have become essential tools for targeted marketing, customer segmentation, and campaign optimization [1,2,3]. These predictive systems enable businesses to personalize outreach, anticipate consumer needs, and allocate resources more efficiently. However, as organizations increasingly rely on automated decision-making, concerns about the fairness of these models have grown—particularly when they affect access to offers, services, or economic opportunities [4,5].

Real-world instances illustrate the stakes of unfair targeting: for example, automated ad-delivery systems have been shown to skew opportunities based on gender or age, potentially violating anti-discrimination norms. In 2022, the U.S. Federal Trade Commission (FTC) issued guidance on deceptive marketing practices and dark patterns, underscoring the need for transparency and fairness in algorithmic systems [6]. These developments highlight how marketing bias is not merely a technical issue, but one with social and regulatory consequences.

Algorithmic bias, whether due to historical data imbalances or unintended correlations in feature space, can result in discriminatory outcomes against certain demographic groups [7,8]. In marketing, such bias may lead to the exclusion or misclassification of high-potential customers based solely on age, education, or socio-economic background, undermining both ethical standards and business performance. This tension between performance and fairness has catalyzed a wave of research and regulatory scrutiny aimed at embedding ethical considerations into machine learning pipelines [6,9,10].

In this study, we investigate how bias manifests in customer targeting models and evaluate practical techniques for detecting and mitigating these disparities. Beyond the technical implementation of fairness, this study draws on broader ethical principles rooted in distributive justice and equal access to opportunity. These theoretical considerations are increasingly relevant in AI governance and motivate the need to assess demographic disparities in model outcomes. By situating fairness within both business and ethical frameworks, we respond to recent calls for more accountable and human-centered AI in commercial contexts.

1.1. Related Work

The intersection of fairness and predictive modeling has become a central topic in machine learning research. Foundational work has shown that discrimination in model outcomes often stems from systemic imbalances in historical data, biased feature selection, or uncalibrated learning objectives [7,8,11]. These concerns have gained practical urgency as ML systems are increasingly deployed in sensitive domains, including finance, hiring, healthcare, and targeted advertising [11,12].

In marketing, fairness has historically received less attention compared to performance optimization. Traditional studies using the Bank Marketing dataset, such as by Moro et al. [13], primarily focused on improving predictive accuracy through feature engineering and advanced classifiers. These efforts, while effective in enhancing AUC and precision, often overlooked demographic fairness across subgroups—a gap this study seeks to address.

Recent surveys have reviewed fairness definitions and auditing techniques. For instance, Verma and Rubin [14] distinguish between group fairness and individual fairness. Garg et al. [15] offer a taxonomy of fairness approaches and challenges. Corbett-Davies and Goel [16] explore legal and ethical trade-offs across parity-based metrics. Toolkits such as AIF360 and Fairlearn have made fairness-aware modeling more accessible, yet their adoption in applied business contexts like marketing remains limited [17,18]. Holstein et al. [19] noted that vague fairness definitions, unclear ROI, and missing operational workflows hinder industry uptake.

Explainability tools such as SHAP (SHapley Additive exPlanations) have become popular for interpreting model behavior and identifying proxies for sensitive attributes [20,21]. While SHAP has seen widespread adoption in healthcare and finance [22,23], its use in customer targeting is still underexplored. Scholars such as Binns [24] argue that fairness in personalization must go beyond statistical parity and consider broader ethical concerns—particularly when exclusion affects access to economic opportunity.

Prior empirical work in other applied domains has examined similar fairness auditing challenges. Gardner et al. [11] evaluated predictive student performance models through slicing analysis and found notable subgroup disparities, underscoring the need for targeted mitigation. In finance, Moro et al. [13] optimized logistic regression for bank telemarketing success, achieving high predictive accuracy but without assessing fairness metrics. Franklin et al. [25] documented sociodemographic biases in biomedical machine learning, illustrating how variables such as age and education can act as proxies for protected attributes—a phenomenon we also identify in marketing data. Hufthammer et al. [17] conducted a systematic comparison of bias mitigation techniques using the AIF360 toolkit, finding that reweighing frequently offered the most favorable balance between fairness and predictive accuracy across multiple datasets. This aligns closely with our empirical results in the marketing domain, reinforcing the robustness of reweighing as a practical intervention. Johnson and Brun [18] introduced Fairkit-learn to evaluate and compare fairness across models, yet their framework did not integrate explainability tools such as SHAP for proxy detection, which we employ to identify potential proxy features.

Audits of real-world ML systems have demonstrated how predictive models can inherit or amplify existing social inequities, especially when variables like age, education, or occupation act as proxies for protected characteristics [25,26]. Our study builds on this literature by applying a full fairness audit to predictive marketing models and evaluating three mitigation strategies—reweighing, threshold adjustment, and feature exclusion—within a replicable framework designed for enterprise deployment.

By addressing limitations of prior work and operationalizing fairness within a marketing context, this paper helps bridge the gap between ethical AI principles and business deployment.

1.2. Research Objectives and Contributions

This study aims to address two central questions:

• To what extent do predictive marketing models exhibit demographic bias in their outputs?
• What mitigation strategies can effectively reduce such bias while maintaining business-relevant performance?

To answer these questions, we develop and evaluate a fairness auditing framework applied to machine learning models trained on the Bank Marketing dataset—a real-world benchmark in customer targeting. We audit two widely used classifiers, logistic regression and random forest, for fairness across sensitive attributes such as age, education level, and job type. Using explainability tools and established fairness metrics like disparate impact ratio and true positive rate parity, we quantify the magnitude and sources of demographic disparities.

We then implement and compare three mitigation techniques: reweighing of training data, group-specific threshold adjustment, and exclusion of sensitive features. Our evaluation considers not only fairness improvements but also changes in accuracy, precision, and AUC, ensuring that fairness gains are not achieved at the expense of business value.

Research Question 1 is addressed through subgroup performance analysis and fairness metrics (DIR, TPR parity, demographic parity difference), while Question 2 is investigated through a comparative evaluation of three mitigation techniques—reweighing, threshold adjustment, and feature exclusion.

This work contributes a reproducible and actionable workflow for fairness auditing in marketing analytics, offering practical insights for enterprise deployment. It also highlights the strategic potential of responsible AI, demonstrating how fairness-aware modeling can enhance inclusion, compliance, and brand equity.

2. Materials and Methods

In this study, we use the Bank Marketing dataset, a publicly available benchmark from the UCI Machine Learning Repository [13], which contains data collected from direct marketing campaigns of a Portuguese banking institution. The goal of the marketing campaign was to predict whether a client would subscribe to a term deposit (binary target variable: yes or no). The Bank Marketing dataset was selected for this study due to its widespread use as a benchmark in predictive modeling, its representation of real-world direct marketing practices, and its inclusion of diverse socio-economic and demographic variables. These characteristics make it a valuable testbed for evaluating both algorithmic performance and fairness across sensitive groups such as age, education, and occupation.

The dataset includes 41,188 records and 20 features capturing customer demographics, contact details, campaign-related information, and socio-economic context.

Table 1. Key Features from the Bank Marketing Dataset used in this study.

Feature	Type	Description	Sample Values (% of Dataset)
y (target)	Binary	Term deposit subscription outcome	yes (11%), no (89%)
age	Numeric	Customer’s age in years	Mean ≈ 41 yrs; <30 (13%), 30–60 (70%), >60 (17%)
job	Categorical	Type of job	admin (23%), blue-collar (21%), technician (17%)
education	Categorical	Highest education level	secondary (51%), tertiary (29%), primary (15%), unknown (5%)
marital	Categorical	Marital status	married (60%), single (30%), divorced (10%)
contact	Categorical	Contact method used in campaign	cellular (65%), telephone (35%)
duration	Numeric	Duration of last contact (in seconds); included for benchmarking only	Mean = 258 s; range = 0–4918; highly right-skewed distribution
outcome	Categorical	Outcome of previous marketing campaign	nonexistent (82%), failure (10%), success (8%)
cons.conf.idx	Numeric	Consumer confidence index (economic indicator)	Range: −50.8 to −26.9

summarizes the main attributes from the Bank Marketing dataset used for modeling and fairness analysis. These features allow us to evaluate both predictive performance and fairness across sensitive attributes, particularly age, education, and job.

Prior to model training, we performed a series of preprocessing steps to prepare the dataset. Categorical variables were encoded using one-hot encoding to retain full granularity while ensuring compatibility with logistic regression and random forest algorithms. Numerical features such as age, duration, and the consumer confidence index were standardized using z-score normalization to improve model stability, especially for logistic regression. We verified that there were no missing values in the dataset, as all rows were complete. Additionally, we applied stratified sampling to split the data into 70% training and 30% test sets, preserving the original class distribution of the target variable (y). These preprocessing steps were applied consistently across all experiments to ensure fairness in model evaluation and comparison.

2.1. Model Training and Evaluation

To model customer subscription behavior, we trained two supervised learning classifiers: logistic regression and random forest. Logistic regression and random forest were selected due to their widespread adoption in enterprise analytics and their contrasting properties [27]. Logistic regression provides a transparent and interpretable linear baseline, making it easier to audit for fairness and explain decisions to stakeholders [27]. Random forest, on the other hand, offers greater modeling capacity through ensemble learning and helps capture non-linear interactions common in behavioral data [28,29]. While more complex models like XGBoost or neural networks may improve raw predictive performance, they often act as black boxes and complicate fairness interpretation and mitigation [30]. Moreover, our focus was to assess fairness in models that are practical, auditable, and already prevalent in marketing deployments. In addition to these two models, we also experimented with support vector machines (SVM) and XGBoost classifiers. However, their performance and fairness outcomes were broadly consistent with those of logistic regression and random forest, and thus were excluded from the final analysis to maintain focus and clarity.

Both models were trained on the preprocessed dataset using 5-fold cross-validation on the training set to tune hyperparameters and assess robustness. For logistic regression, we applied L2 regularization to avoid overfitting. The random forest model was trained with 100 estimators and default parameters for maximum depth and minimum samples per split, with further tuning based on grid search. The classification threshold was set at 0.5 for all baseline evaluations. Protected group definitions followed demographic groupings outlined in AIF360 documentation, and the same definitions were applied consistently across all experiments.

Model performance was evaluated on the test set using standard classification metrics, including:

• Accuracy: Overall proportion of correct predictions.
• Precision: Proportion of true positives among predicted positives, relevant to targeting efficiency.
• Recall: Proportion of true positives among actual positives.
• AUC (Area Under the ROC Curve): Overall discriminative ability of the model.

In addition to these global metrics, we computed subgroup-specific performance (precision and recall for different age and education groups) to assess disparities and support fairness analysis.

2.2. Fairness Auditing Framework

To assess fairness in our predictive models, we adopted a multi-metric auditing approach, focusing on both group-level disparities and explainability of model behavior. The fairness audit concentrated on three sensitive attributes: age, education, and job type, which were selected due to their potential impact on both model predictions and ethical risk. We computed the following fairness metrics, all widely used in algorithmic fairness research and regulatory contexts [14,15,26,31,32]:

• Disparate Impact Ratio (DIR): The ratio of positive prediction rates between a protected group and a reference group. A DIR below 0.8 is generally considered indicative of potential bias and is often referenced in compliance guidelines such as the “80% rule” [31].
• True Positive Rate (TPR) Parity: Assesses whether the true positive rates (recall) are equal across demographic subgroups, highlighting disparities in beneficial outcomes [32]. We also report the corresponding TPR Parity Gap (ΔTPR) for consistency with difference-based measures.
• Demographic Parity Difference: The absolute difference in positive prediction rates between groups, regardless of actual outcome; a stricter statistical parity measure that is straightforward to communicate to non-technical stakeholders [14]. We note that the ratio form (Demographic Parity Ratio, DPR) is conceptually related to DIR and may be reported in supplementary analysis.

These metrics were selected based on their widespread adoption in both academic research and regulatory guidance. While other fairness metrics exist, such as equalized odds or calibration. We prioritized these three due to their interpretability, regulatory significance, and actionable nature in enterprise deployments.

To interpret the model’s decision logic, we used SHAP (SHapley Additive exPlanations), a model-agnostic technique that quantifies each feature’s contribution to individual predictions. This enabled us to detect proxy variables that might indirectly encode sensitive information, as well as to visualize and compare the importance of features across models. Fairness metrics were reported both before and after applying bias mitigation strategies for clear picture of model behavior and fairness-aware training regimes.

2.3. Bias Mitigation Techniques

To address the demographic disparities uncovered in the fairness audit, we implemented and evaluated three bias mitigation strategies that target different stages of the machine learning pipeline:

• Reweighing (Pre-processing): Originally proposed by Kamiran and Calders [33], this method adjusts the weights of training samples based on group–label combinations to reduce inherent bias in the data distribution. We used the implementation in IBM’s AIF360 library with default weighting schema and stratified by each protected attribute. This technique was selected due to its demonstrated effectiveness in balancing fairness and performance in prior comparative studies [17,33].
• Threshold Adjustment (Post-processing): Based on Hardt et al.’s equalized odds post-processing framework [32], this approach calibrates group-specific classification thresholds after model training to equalize the TPR across protected groups. Implementation followed Fairlearn’s ThresholdOptimizer with the constraint “equal opportunity,” tuned on the validation set. This method was chosen to address disparities without altering the training process.
• Feature Exclusion (In-processing variant): A common compliance-oriented practice [26,34], in which sensitive attributes (age, education) were excluded prior to training to evaluate whether performance could be retained without direct use of potentially biased predictors. Models were retrained using identical hyperparameters to the baseline, and SHAP analysis was used to check for proxy effects.

We selected these techniques to represent three intervention points—pre-processing, post-processing, and feature-level adjustment—ensuring coverage of commonly used approaches with minimal deployment complexity. Other methods, such as adversarial debiasing [35] or reject option classification [33], were considered but not implemented due to their higher computational cost and integration complexity in the targeted business intelligence environment.

3. Results

3.1. Baseline Model Performance

We began our analysis by training and evaluating two baseline models: logistic regression and random forest. These served as benchmarks for assessing both predictive accuracy and fairness prior to the application of any mitigation techniques. On the test set, the logistic regression model achieved an accuracy of 0.89, with a precision of 0.72 and AUC of 0.85, while the random forest model performed slightly better, with an accuracy of 0.91, precision of 0.75, and AUC of 0.88. These results highlight the predictive power of ensemble models in capturing complex, non-linear relationships within customer behavior data. While the random forest model offered higher predictive accuracy, subgroup analysis revealed that this gain did not necessarily translate into equitable performance across demographic groups. These findings reinforce the concern that strong overall accuracy can mask underlying biases in subgroup-level predictions, which is critical in contexts such as marketing where access to offers must be fairly distributed.

The comparative Receiver Operating Characteristic (ROC) curves in Figure 1 further illustrate this difference in discriminative ability, with the random forest model consistently achieving higher true positive rates across most false positive thresholds.

Despite strong overall performance, subgroup analysis revealed demographic disparities. As shown in

Table 2. Subgroup-wise performance metrics across age and education levels for both models (mean ± SE from 5-fold CV).

Demographic Group	TPR (Logistic Regression)	Precision (Logistic Regression)	TPR (Random Forest)	Precision (Random Forest)
Age: 18–30	0.78 ± 0.02	0.70 ± 0.01	0.81 ± 0.02	0.73 ± 0.01
Age: 31–60	0.74 ± 0.01	0.73 ± 0.01	0.77 ± 0.01	0.76 ± 0.01
Age: 60+	0.62 ± 0.02	0.66 ± 0.02	0.65 ± 0.02	0.68 ± 0.02
Education: Primary	0.65 ± 0.02	0.67 ± 0.01	0.69 ± 0.02	0.70 ± 0.01
Education: Secondary	0.72 ± 0.01	0.71 ± 0.01	0.76 ± 0.01	0.74 ± 0.01
Education: Tertiary	0.79 ± 0.01	0.76 ± 0.01	0.81 ± 0.01	0.78 ± 0.01

, the true positive rate (TPR) for older individuals (age 60+) was notably lower than for those aged 30–40, particularly in the logistic regression model. Similarly, customers with only primary education had lower TPRs and precision compared to those with tertiary education, suggesting that the models were less effective for certain groups—even when actual subscription behavior was similar.

These early findings indicate that demographic bias is embedded in the baseline models, warranting a deeper fairness audit in subsequent sections.

3.2. Fairness Auditing Results

Following the evaluation of baseline performance, we conducted a comprehensive fairness audit across three sensitive demographic attributes: age, education, and job type. Our goal was to determine whether the predictive models demonstrated systematic disparities in outcomes that could result in unfair or discriminatory targeting.

We computed a set of standard fairness metrics for both the logistic regression and random forest models:

• Disparate Impact Ratio (DIR): The DIR for older individuals (age 60+) was 0.65 in the logistic regression model and 0.68 in the random forest model. Both fall below the commonly accepted threshold of 0.80, signaling potential bias in how positive outcomes are distributed across age groups.
• True Positive Rate (TPR) Parity: The TPR gap between individuals with primary and tertiary education exceeded 0.14 for logistic regression, indicating a skewed allocation of correct positive predictions favoring higher-educated individuals.
• Demographic Parity Difference: We observed that certain occupational groups, such as “blue-collar” workers, received positive predictions at a significantly lower rate than others like “management,” suggesting the influence of indirectly biased features.

To investigate the sources of these disparities, we utilized SHAP (SHapley Additive exPlanations) to interpret model behavior. SHAP analysis played a key role in not only interpreting feature importance but also in uncovering potential proxies for sensitive features. For example, “job_blue-collar” and “education_primary” consistently emerged as top predictors, indicating that even if direct features are removed, proxy-based bias may persist in model behavior. This finding further justifies the need for explanation-based fairness auditing. As shown in Figure 2, features such as age, job_blue-collar, and education_primary ranked among the most influential predictors in the random forest model. Their elevated SHAP values imply not only high predictive power but also potential for encoding bias, especially when these features correlate with protected characteristics.

The computed fairness metrics, summarized in

Table 3. Fairness metrics across sensitive groups for logistic regression and random forest models (mean ± SE).

Group Comparison	Disparate Impact (Logistic)	TPR Parity Gap (Logistic)	Demographic Parity Diff (Logistic)	Disparate Impact (RF)	TPR Parity Gap (RF)	Demographic Parity Diff (RF)
Age: 60+ vs. 30–60	0.65 ± 0.02	0.13 ± 0.01	0.13 ± 0.01	0.68 ± 0.02	0.14 ± 0.01	0.11 ± 0.01
Education: Primary vs. Tertiary	0.71 ± 0.02	0.12 ± 0.01	0.11 ± 0.01	0.74 ± 0.02	0.12 ± 0.01	0.09 ± 0.01
Job: Blue-collar vs. Management	0.68 ± 0.02	0.10 ± 0.01	0.10 ± 0.01	0.67 ± 0.02	0.10 ± 0.01	0.08 ± 0.01

, further validate that both models exhibit demographic biases—particularly along age and education dimensions—which could lead to inequitable marketing outcomes if left unaddressed.

These results confirm the presence of algorithmic bias in both models and emphasize the need for systematic mitigation before these tools can be deployed responsibly in enterprise marketing pipelines. The SHAP summary plot (Figure 2) reveals the top features influencing model predictions, highlighting age, job, and education as key contributors to potential bias. The fairness metrics table (Table 3) quantifies disparities across models, showing that both logistic regression and random forest exhibit demographic biases, particularly in age and education.

For completeness, we also trained Support Vector Machine (SVM) and XGBoost classifiers using the same preprocessing, evaluation metrics, and subgroup analyses. Results for SVM and XGBoost (Appendix A,

Table A1. Subgroup-wise performance metrics across age and education levels for SVM and XGBoost models (mean ± SE from 5-fold CV).

Demographic Group	TPR (SVM)	Precision (SVM)	TPR (XG Boost)	Precision (XG Boost)
Age: 18–30	0.77 ± 0.02	0.71 ± 0.01	0.80 ± 0.02	0.73 ± 0.01
Age: 31–60	0.73 ± 0.01	0.74 ± 0.01	0.76 ± 0.01	0.76 ± 0.01
Age: 60+	0.63 ± 0.02	0.67 ± 0.02	0.66 ± 0.02	0.69 ± 0.02
Education: Primary	0.66 ± 0.02	0.68 ± 0.01	0.70 ± 0.02	0.71 ± 0.01
Education: Secondary	0.71 ± 0.01	0.72 ± 0.01	0.75 ± 0.01	0.74 ± 0.01
Education: Tertiary	0.78 ± 0.01	0.76 ± 0.01	0.81 ± 0.01	0.78 ± 0.01

and

Table A2. Fairness metrics across sensitive groups for SVM and XGBoost models (mean ± SE).

Group Comparison	Disparate Impact (SVM)	TPR Parity Gap (SVM)	Demographic Parity Diff (SVM)	Disparate Impact (XGBoost)	TPR Parity Gap (XGBoost)	Demographic Parity Diff (XGBoost)
Age: 60+ vs. 30–60	0.66 ± 0.02	0.12 ± 0.01	0.11 ± 0.01	0.67 ± 0.02	0.13 ± 0.01	0.10 ± 0.01
Education: Primary vs. Tertiary	0.72 ± 0.02	0.12 ± 0.01	0.10 ± 0.01	0.73 ± 0.02	0.12 ± 0.01	0.09 ± 0.01
Job: Blue-collar vs. Management	0.69 ± 0.02	0.10 ± 0.01	0.09 ± 0.01	0.68 ± 0.02	0.10 ± 0.01	0.08 ± 0.01

) closely mirror those of logistic regression and random forest in both performance and fairness metrics. Minor numerical differences (typically within ±0.02) did not alter the observed bias patterns, indicating that the disparities are model-agnostic. Therefore, we focus on logistic regression and random forest in the main text for clarity and brevity.

3.3. Mitigation Outcomes

To address the demographic disparities identified in our fairness audit, we applied three bias mitigation techniques—reweighing, threshold adjustment, and feature exclusion—and evaluated their effects on both fairness and model performance. Each technique targeted a different stage of the machine learning pipeline and offered unique trade-offs.

Reweighing (Pre-processing) yielded the most promising results. After adjusting instance weights during training to balance group-label distributions, we observed an improvement in fairness metrics across the board. The disparate impact ratio for the 60+ age group improved from 0.65 to 0.82, and true positive rate parity gaps decreased by up to 40% relative to the baseline. Importantly, accuracy and precision experienced only marginal declines (typically < 2%), demonstrating that fairness gains were not achieved at the expense of business relevance.

Threshold Adjustment (Post-processing) was effective in reducing TPR parity gaps across groups by customizing decision thresholds per demographic category. While this method improved fairness—raising TPR parity scores to acceptable levels—it required additional logic for deployment and careful calibration to avoid overfitting thresholds to historical patterns.

Feature Exclusion (In-processing) involved retraining models without sensitive attributes such as age and education. Although this approach led to slight improvements in disparate impact ratios, it caused a noticeable drop in precision and AUC (by approximately 4–6%), indicating that these features were highly predictive. Furthermore, SHAP analysis suggested that proxy variables (e.g., job type or contact method) continued to encode similar information, reducing the overall effectiveness of exclusion.

As illustrated in Figure 3, the bar chart highlights changes in fairness and performance metrics across all three strategies.

Table 4. Summary table of metric changes for each mitigation strategy.

Metric	Baseline	Reweighing	Threshold Adjustment	Feature Exclusion
Accuracy	0.85 ± 0.01	0.83 ± 0.01	0.84 ± 0.01	0.81 ± 0.01
Precision	0.78 ± 0.01	0.76 ± 0.01	0.77 ± 0.01	0.74 ± 0.01
Disparate Impact Ratio	0.65 ± 0.02	0.82 ± 0.02	0.76 ± 0.02	0.71 ± 0.02
TPR Parity	0.16 ± 0.01	0.09 ± 0.01	0.07 ± 0.01	0.12 ± 0.01
Demographic Parity Difference	0.10 ± 0.01	0.06 ± 0.01	0.04 ± 0.01	0.08 ± 0.01

summarizes the raw and normalized values, supporting our conclusion that reweighing offers the most effective trade-off between fairness and accuracy. Accuracy and precision values represent the mean ± standard deviation over five independent runs with different random seeds. Disparate Impact Ratio (DIR), True Positive Rate Parity Difference (ΔTPR), and Demographic Parity Difference (DPD) were computed for the most impacted protected group in each category (age, education, job), then averaged across all categories. DPD was calculated as the absolute difference in positive prediction rates between the protected and reference groups. These findings suggest that reweighing provides a practical balance between fairness and performance, while threshold adjustment may serve as a complementary tool. Feature exclusion, while well-intentioned, may introduce new issues without eliminating bias. These mitigation strategies illustrate the inherent tension between fairness and business-relevant performance. Notably, reweighing yielded the most balanced outcome, but its implementation requires access to group-label distributions and may affect scalability. Threshold adjustment demands careful governance in real-world deployments, while feature exclusion, despite its simplicity, proved insufficient alone—emphasizing that fairness solutions must be context-sensitive rather than one-size-fits-all.

3.4. Comparative Summary and Trade-Offs

Reweighing emerges as the most balanced approach, delivering strong fairness gains without significantly compromising predictive performance. Specifically, it improved the Disparate Impact Ratio to 0.82 and reduced the TPR parity gap by seven percentage points, while preserving overall accuracy (0.83) and precision (0.76). These results position reweighing as a high-leverage, low-friction intervention for enterprises seeking to embed fairness into machine learning systems.

Threshold adjustment, on the other hand, achieved the best results for TPR parity (gap reduced to 0.07), making it useful where recall fairness is prioritized—such as in public-facing or compliance-sensitive applications. However, its deployment complexity, including the need for group-level segmentation and ongoing tuning, may limit its practicality.

Feature exclusion, while conceptually simple and often recommended for compliance purposes, underperformed in this context. Removing sensitive attributes like age and education led to measurable losses in model precision and had limited fairness impact, owing to the presence of proxy variables that indirectly encode the same demographic information. This outcome reinforces the importance of explainability tools (e.g., SHAP) in guiding responsible feature selection and avoiding naïve bias mitigation tactics.

Ultimately, our results demonstrate that no mitigation strategy is universally optimal. The best approach depends on an organization’s business context, ethical priorities, and technical resources. Our framework provides a replicable and data-driven method for evaluating these trade-offs—equipping enterprises to make more informed, fair, and effective use of AI in customer targeting. From our results, reweighing improved the Disparate Impact Ratio by approximately 26% while reducing accuracy by only ~2%, representing the most favorable proportional balance among the tested methods. Threshold adjustment achieved similar fairness gains but with a slightly higher accuracy reduction, while feature exclusion yielded smaller fairness improvements alongside greater performance loss.

Our work differs from existing studies by operationalizing fairness techniques within a real-world enterprise context, offering an implementation-oriented perspective. By balancing accuracy with fairness and evaluating strategy-specific trade-offs, we provide a roadmap for organizations aiming to build ethically robust and practically viable AI systems in marketing.

4. Discussion

The fairness audit conducted in this study revealed consistent disparities in model outcomes across key demographic groups, particularly with regard to age and education level. These results build on and extend prior research in algorithmic fairness and predictive modeling [7,8,14], confirming that bias is not merely a byproduct of model selection but is deeply rooted in data representation and feature correlations. Our findings affirm that traditional accuracy metrics alone are insufficient for assessing the social impact of ML models in applied contexts such as marketing. This work contributes to a growing body of literature that emphasizes fairness as a core model performance dimension. In the baseline models, older individuals and those with basic education were systematically less likely to receive positive predictions, despite having similar response behaviors to other groups. This bias was evident across both logistic regression and random forest classifiers and was quantitatively supported by low disparate impact ratios (e.g., DIR < 0.7) and substantial gaps in true positive rates.

The application of mitigation strategies produced encouraging results. Reweighing, in particular, significantly improved group fairness metrics, elevating DIR values above the commonly accepted 0.8 threshold while maintaining strong accuracy and precision. Threshold adjustment further closed the gap in true positive rates between age groups, while feature exclusion—though conceptually appealing—led to a noticeable loss in model performance, highlighting the trade-offs of excluding informative yet sensitive features. These findings suggest that fairness-aware preprocessing methods like reweighing can be powerful tools for reducing bias without compromising business value.

4.1. Ethical and Regulatory Implications

The presence of algorithmic bias in predictive marketing models raises important ethical and legal concerns. Discriminatory targeting—whether intentional or the byproduct of historical patterns in training data—can reinforce existing inequalities and expose organizations to reputational and regulatory risks. In regions governed by the General Data Protection Regulation (GDPR), such disparities may trigger scrutiny under principles of fairness, transparency, and the right to explanation. Similarly, in the U.S., the Federal Trade Commission (FTC) has issued guidance emphasizing the need for fairness, accountability, and transparency in AI-driven decision systems.

Our findings support the growing consensus that fairness is not merely a technical add-on but a core requirement for responsible AI. By demonstrating how relatively simple interventions such as reweighing and threshold adjustment can measurably improve fairness outcomes, we show that ethical compliance is both feasible and practical. Moreover, using explainability tools like SHAP to uncover the influence of individual features adds a layer of interpretability that can assist in legal defensibility and internal auditing. As regulators increasingly demand transparency and fairness in automated systems, integrating such auditing mechanisms early in the modeling pipeline will be essential for long-term sustainability and compliance.

4.2. Strategic Business Value of Responsible AI

Beyond regulatory compliance and ethical obligation, fairness in predictive marketing presents a compelling business case. Biased models not only risk excluding valuable customer segments but also erode consumer trust—an increasingly vital currency in the digital economy. Responsible AI practices help companies avoid reputational damage while expanding market reach by ensuring that marketing campaigns equitably include a wider range of demographics.

In our study, mitigation techniques like reweighing maintained strong predictive performance while significantly improving fairness metrics. This demonstrates that inclusive models do not necessarily come at the expense of business efficiency. In fact, fairer models may enhance the long-term return on investment by identifying underserved but responsive customer groups, ultimately leading to more effective and sustainable marketing strategies.

Moreover, transparency and fairness are becoming key differentiators in brand positioning. Enterprises that adopt fairness-aware analytics workflows signal their commitment to ethical standards, which can strengthen stakeholder relationships and customer loyalty. As fairness becomes integrated into enterprise business intelligence pipelines, it evolves from a compliance checkbox into a strategic asset—enabling innovation, resilience, and reputational strength in a competitive landscape. From a business intelligence (BI) perspective, integrating fairness audits directly into customer segmentation and campaign optimization processes enables marketing teams to better align outreach strategies with DEI (diversity, equity, and inclusion) initiatives. This study provides an example of how organizations can practically embed fairness checkpoints into existing analytics workflows without significant disruption.

4.3. Comparison with Prior Studies

The fairness improvements observed in this study align with and extend findings from related work in other applied domains. Hufthammer et al. [17] found that reweighing frequently offered the most favorable balance between fairness and accuracy in classification tasks; our results similarly show that reweighing increased the Disparate Impact Ratio for older individuals from 0.65 to 0.82, with only a 0.02 decrease in precision. Johnson and Brun [18] emphasized the value of comparative fairness evaluation across models, a principle we operationalized by benchmarking three mitigation strategies in parallel. In contrast, Moro et al. [13] and Ghatasheh [28] focused on optimizing predictive accuracy in marketing and credit risk contexts, respectively, but did not incorporate fairness metrics—our findings demonstrate that high predictive performance can be maintained while addressing equity concerns. Franklin et al. [25] documented how socio-demographic features can act as proxies for protected characteristics in biomedical ML models; our SHAP analysis confirmed similar patterns in marketing, with occupation and education serving as indirect indicators of age groups.

Differences in effect sizes and performance–fairness trade-offs across studies are likely due to variations in dataset composition, domain-specific target distributions, and the sequencing of mitigation strategies. By integrating explainability-based proxy detection with multiple fairness interventions, our work bridges the gap between theoretical fairness principles and practical, domain-specific deployment in predictive marketing.

4.4. Limitations

While our findings offer important insights, several limitations must be acknowledged. First, the analysis was conducted on a single dataset—the Bank Marketing dataset—which, while rich and widely used, may not fully capture the complexity and diversity of customer behavior across industries or geographies. The generalizability of our results to other domains (e.g., retail, insurance, or healthcare marketing) requires further validation.

Second, group membership in our fairness audit was based on explicit demographic variables such as age and education. In many real-world applications, these attributes may not be directly available due to privacy concerns, necessitating the use of proxy features or inferred group membership, which can introduce noise or unintended bias into the audit process.

Third, our models operated in a static batch-processing pipeline, meaning that they did not account for time-sensitive or real-time dynamics that are often present in production systems. This limits the applicability of our findings to streaming or online-learning environments, where bias may evolve over time and require continuous monitoring.

Finally, while we explored three common mitigation strategies, a broader exploration of hybrid or adversarial methods may reveal more effective solutions. These limitations suggest important directions for future research aimed at creating more robust and adaptive fairness frameworks. Additionally, our framework does not yet fully integrate theoretical constructs from behavioral economics or customer psychology, which could offer richer insights into the reasons behind differential model behavior across demographic groups. Bridging this gap between fairness theory and marketing practice is an important area for future research, especially in contexts where personalization intersects with perceived consumer autonomy.

4.5. Recommendations for Deployment in Enterprise BI Tools

The implementation of fairness-aware pipelines in business intelligence (BI) systems must reflect both operational constraints and ethical principles grounded in fairness theory. Concepts like accountability, transparency, and stakeholder-inclusive design discussed in the fairness-aware system design literature are essential to guide effective deployment strategies.

Translating fairness auditing and mitigation into enterprise business intelligence (BI) environments requires both technical integration and organizational alignment. Based on our findings, we recommend a modular workflow that begins with fairness-aware preprocessing (e.g., reweighing) and includes regular performance audits across demographic segments. This approach ensures that equity is embedded from the outset and monitored throughout the model lifecycle.

Enterprises should incorporate fairness dashboards alongside traditional BI metrics, allowing decision-makers to balance model accuracy with fairness KPIs in real time. Tools like SHAP can be embedded within model monitoring platforms to help analysts interpret feature impacts and detect proxy discrimination. Automation of threshold tuning or reweighing strategies can also be considered to adapt to shifts in customer behavior or demographics over time.

Critically, fairness considerations must be aligned with business objectives. Stakeholders across data science, compliance, marketing, and legal teams should collaborate to define fairness goals, assess trade-offs, and validate deployment policies. Education and documentation are essential—translating complex model behavior into understandable terms for executive review helps build trust in fairness-aware systems.

By embedding responsible AI practices into BI toolchains, organizations can not only mitigate legal and reputational risks but also unlock competitive advantages through inclusive and transparent customer engagement strategies.

5. Conclusions

This study underscores the critical importance of integrating fairness auditing and mitigation into predictive marketing workflows. Using the Bank Marketing dataset, we trained and evaluated logistic regression and random forest models to predict customer subscription behavior, uncovering significant disparities in outcomes across demographic groups—particularly age, education, and job type. Our fairness audit, employing metrics such as the disparate impact ratio and true positive rate parity, revealed that these models, if deployed without scrutiny, could result in biased targeting that not only risks ethical and legal concerns but also undermines marketing effectiveness.

To address these issues, we implemented and compared three mitigation strategies: reweighing, threshold adjustment, and feature exclusion. Among these, reweighing offered the best balance between improving fairness and preserving performance metrics such as accuracy and precision. Threshold adjustment was effective but introduced operational complexity, while feature exclusion resulted in a notable loss of predictive power and incomplete mitigation due to proxy variables. These findings demonstrate that fairness-aware modeling is both feasible and beneficial in real-world business contexts.

Moreover, the use of SHAP explainability tools allowed us to trace potential bias to specific features, enhancing interpretability and supporting transparency—a key requirement for regulatory compliance. Our proposed workflow provides a replicable and practical framework that enterprise teams can adopt to embed fairness directly into their machine learning pipelines.

As fairness becomes a strategic priority for both ethical governance and competitive differentiation, organizations must evolve from reactive compliance to proactive responsibility in AI deployment. Future work can extend this research to dynamic and real-time marketing systems, multiple datasets, and additional bias mitigation methods. Ultimately, the adoption of fairness-aware analytics is not just about reducing harm—it is about building trust, expanding opportunity, and enabling more inclusive and accountable innovation in the digital economy.