Utilizing data sampling techniques on algorithmic fairness for customer churn prediction with data imbalance problems

Area of Science:

• Machine Learning
• Data Science
• Algorithmic Fairness

Background:

• Customer churn prediction (CCP) is crucial for service providers, but often suffers from class imbalance problems (CIP).
• Data sampling techniques (DSTs) are used to mitigate CIP, but their impact on algorithmic fairness is not well understood.
• Algorithmic fairness, particularly regarding gender discrimination, is an increasingly important consideration in machine learning applications.

Purpose of the Study:

• To investigate the effect of DSTs on algorithmic fairness in CCP.
• To compare the performance and fairness of classification models before and after applying DSTs.
• To identify potential gender-based discrimination introduced or exacerbated by DSTs.

Main Methods:

• Reviewed four common DSTs applied to three real-world imbalanced datasets.
• Utilized six popular classification techniques for CCP.
• Evaluated both classifier performance and algorithmic fairness using established metrics, focusing on gender disparities.

Main Results:

• Random Forest demonstrated superior performance across all datasets.
• SMOTE and ADASYN techniques were found to increase discrimination against the female group.
• Unintentional discrimination was higher in original imbalanced data for Logistic Regression, LightGBM, and XGBoost.

Conclusions:

• There is a significant gap in systematic research on DSTs' impact on algorithmic fairness in CCP.
• This study provides critical insights into how sampling methods can affect fairness in churn prediction models.
• Findings emphasize the need to consider algorithmic fairness when applying DSTs in real-world CCP scenarios.