An unbalanced data classification method based on improved SMOTE
Fangrui Shi, Guochao Fan · 2023
With the continuous development of modern industry, product quality control standards have become increasingly stringent. The classification of product quality characteristics is of paramount importance to product quality control. It is necessary to improve classification methods specifically for imbalanced data to solve the issue of data imbalance in the classification of product quality characteristics. The Synthetic Minority Over-Sampling Technique (SMOTE) is a mature algorithm capable of effectively classifying imbalanced data and is widely applied to various classification problems. However, its performance in handling class overlap and noise issues leaves room for improvement, and these issues are often prevalent in the classification of product quality characteristics. This paper proposes an improved data sampling classification algorithm based on KNN under-sampling and DBSCAN methods to enhance the SMOTE algorithm. We first employ the KNN algorithm to remove part of the majority class data in class overlap regions, then introduce DBSCAN to eliminate noise within the sample data. Lastly, we apply SMOTE for over-sampling to balance the data. We compare common imbalanced data classification methods and our improved KD-SMOTE algorithm using a decision tree classifier on five groups of public UCI datasets. Experimental results indicate that our method improves AUC and other metrics, providing a significant enhancement over existing algorithms. This method better caters to the needs of product quality characteristic classification.