UFIDSF: An undersampling approach based on feature importance and double side filter for imbalanced data classification

Abstract

Imbalanced data has the potential to detrimentally impact the efficacy of machine learning algorithms. If imbalanced data is not effectively processed, it will have a great impact on the classification results and reduce the reliability and practicability of modeling, so it has received widespread attention. From the past few decades to the present, various methods have emerged to solve the problem of imbalance data classification. The most common method is to start from the data level and realize data balance by resampling method. However, it remains a challenge to ensure that more valuable data is learned during the resampling process. Therefore, this study proposes an undersampling framework (UFIDSF) based on feature importance and double side filter. The first novelty of this framework is the use of double side filter to filter noise data in both majority and minority class samples. The second novelty is the projection of data samples into one dimension. UFIDSF is realized by calculating the distance between the feature of each dimension of the sample and its nearest neighbor and combining the feature importance. Experiments were conducted on 30 common benchmark imbalanced datasets, comparing the performance of 10 methods across four classifiers. Experimental results show that the proposed UFIDSF is effective and stable, and can significantly improve the adverse effects of machine learning algorithms on imbalanced data.