Exploring imbalanced data challenges: oversampling efficacy and sample size estimation

Abstract

Class imbalance affects the accuracy and generalization of predictive models, making it essential to explore efficient strategies to mitigate this issue. In this thesis, we investigate the effectiveness of oversampling techniques and propose a method for estimating the optimal sample size in imbalanced classification problems. The results indicate that optimizing the decision threshold can replace the need for synthetic data generation, reducing reliance on oversampling. Additionally, the proposed methodology allows for the estimation of the necessary sample size to ensure more stable classifications, avoiding excessive data collection. Thus, this research contributes to understanding the impact of balancing techniques and provides more efficient alternatives for improving model performance. The proposed approach enables more informed decisions regarding sampling and preprocessing, minimizing the need for artificial data manipulation.