A New Safe-Level Enabled Borderline-SMOTE for Condition Recognition of Imbalanced Dataset
Chao Chen, Wei Shen, Chenhao Yang, Wei Fan, Xin Liu, Ying Li
Abstract
Machine learning-based classification strategy has been successfully applied in actual industrial monitoring but it is often hindered when the data set is imbalanced. Technically, the misclassification phenomenon, as a serious performance degradation of generalisation ability, often occurs in minority class. For this problem, Borderline-Synthetic Minority Over-sampling Technique (B-SMOTE), which aims to enrich the quantity of minority samples around decision boundaries, has received considerable attention. However, most imbalanced classification techniques under the framework of B-SMOTE generate instances by a random weight number from 0 to 1, which may result in an authentic reduction of newly-born samples. Herein, a novel over-sampling strategy, which aims to provide new safety criteria and reassign the threshold of weight coefficient, is proposed to boost the authenticity of generated samples and classification accuracy. In addition, light gradient boosting machine (LightGBM) is adopted to build the classification model. Related experiments show the effectiveness and superiority of the proposed method in handling imbalanced classification tasks.