Oversampling and imputation for imbalanced missing data

Abstract Oversampling and imputation for imbalanced missing data Imbalanced data is a widespread issue that is naturally occurring. For instance, fraudulent banking transactions are less frequent than normal transactions, and the number of cancer cases is lower than that of patients. Furthermore, it is common for data to contain missing values. Dealing with imbalanced and missing data can present several challenges during analysis, potentially leading to biased models or skewed performance evaluation. Therefore, it is important to recognize these issues and address them in the data analysis process as needed. However, imbalance and missingness are usually handled separately, which can be inefficient and increase computational costs. We have developed a new imputation-based sampling method called OsImp, designed specifically for imbalanced datasets with high rates of missing data. This method, suitable for small and large multi-class datasets, can handle categorical, continuous, or mixed data as long as an appropriate imputation algorithm is used. It’s also worth noting that OsImp can be applied to datasets without missing data and still perform competitively compared to state-of-the-art methods for imbalanced data. Our initial experiments have shown that this method outperforms many current approaches across a variety of datasets.