Mitigating class imbalance in churn prediction with ensemble methods and SMOTE

Area of Science:

• Machine Learning
• Data Science
• Predictive Analytics

Background:

• Imbalanced datasets pose a significant challenge in machine learning, leading to biased models.
• This bias particularly affects predictive analytics applications like customer churn prediction.
• Accurate prediction requires addressing data skewness.

Purpose of the Study:

• To investigate the impact of imbalanced datasets on machine learning model accuracy.
• To evaluate the performance of various classifiers and ensemble models on imbalanced data.
• To assess the effectiveness of data balancing techniques in improving predictive performance.

Main Methods:

• Utilized a churn dataset to evaluate model performance under data imbalance.
• Compared nine individual classifiers and six homogeneous ensemble models.
• Applied the SMOTE (Synthetic Minority Over-sampling Technique) sampling method to balance the dataset.

Main Results:

• Ensemble models showed improved predictive performance over single classifiers on imbalanced data, though accuracy remained suboptimal.
• Post-SMOTE balancing, model performance increased from 61% to 79%.
• Adaboost achieved the highest performance with an F1-Score of 87.6% in churn prediction.

Conclusions:

• Data balancing is crucial for accurate machine learning predictions.
• SMOTE effectively mitigates bias in imbalanced datasets.
• Adaboost demonstrates superior performance for churn prediction tasks.