DPB-SMOTE: Discrete Probability-Based SMOTE for Classification Process (Case Study: Jeruk Pontianak)

Training a machine learning model with imbalanced data might introduce bias, significantly reducing the model's ability to generate precise predictions. Thus, ensuring sample balance is crucial. This paper proposes a novel synthetic minority over-sampling technique (SMOTE) for datasets with discrete features, called the discrete probabilitybased SMOTE or DPB-SMOTE. The DPB-SMOTE is an improvement to the SMOTE algorithm which incorporates the probability theory and borderline samples into the SMOTE. Based on numerical simulations conducted on 30 randomized jeruk Pontianak (citrus nobilis var. microcarpa) datasets, the DPB-SMOTE, when combined with a classification tree (DPBSMOTE + CT) achieved higher model predictive accuracy than SMOTE + CT. Furthermore, the proposed method attained an AUC of 0.940, exceeding the AUC of 0.911 achieved by SMOTE + CT. These results indicate that the DPB-SMOTE + CT outperforms the SMOTE + CT in classifying positive class.