Synthetic-MixUp: A Simple Framework for Imbalanced Text classification

This study focuses on analyzing effective strategies to handle imbalanced data in text classification. The study employs probabilistic Naive Bayes as a common machine learning approach and investigates two oversampling techniques: Random Oversampling (ROS) [4] and Synthetic Minority Oversampling Technique (SMOTE) [2]. The methodology comprises several stages, starting with preprocessing the data into vector representation using TF-IDF from document samples. To address the imbalanced data, we select some vectors from a single document and mix them with other documents to generate a new vector representation. Finally, the study employs the Naive Bayes (NB) [3] classification method. The experimental results confirm the efficacy of the proposed methods, showcasing notable enhancements in the performance of minority classes. By successfully addressing the challenge of class imbalance, our approach enables the development of accurate text classification models, Which can be observed from both NB+ROS and NB+SMOTE generally outperform the NB+Imbalanced Data that F1-Score, Average Precision and Balanced Accuracy around 83.65; 83.65; 85.8. From our approach, NB+SMOTE tends to yield better results compared to NB+ROS in most cases.