Advanced Weighted Approach for Class Imbalance Learning

Abstract

Predictive models derived from statistical learning techniques often assume that data originate from simple random sampling, thus assigning equal weight to all individuals. However, this assumption faces two significant challenges: it overlooks the complexity of real samples, where individuals may have different sampling weights, and it introduces a bias toward the majority class in imbalanced datasets. In this study, we propose an innovative approach that introduces differentiated weights for individuals by adjusting sample weights through calibration. This method aims to address class imbalance issues while improving the representativeness of samples. We applied it to the Support Vector Machine. Additionally, we developed an improved adjusted weighting approach to further enhance model performance, particularly for the minority class. This improved version combines two widely used techniques for handling class imbalances (resampling and cost-sensitive learning) by first balancing the classes through resampling, then applying adjusted sample weights during training. We evaluated the performance of our approach on real datasets with varying levels of imbalance using multiple evaluation metrics. The results were compared with various conventional methods commonly employed to address class imbalance. Our findings demonstrate the relevance and generalizability of our proposed algorithms, which often achieve performance equal to or better than that of established competing methods. Overall, our methodology not only corrects sample imbalances but also ensures a more accurate representation of the target population in the model, making it a robust and flexible solution for real-world imbalanced classification challenges.