A Study on Imbalanced Data Sentiment Classification of Customer Reviews

As for the imbalanced sentiment classification problem, a Chinese sentiment classification model based on consumer reviews is proposed for three public datasets with different imbalance situations, which can effectively predict the minority class. SMOTE, Borderline-SMOTE, and ADASYN oversampling methods are chosen to deal with the data imbalance problem without sacrificing the majority class. Under the three feature representations, the experiments are compared and analyzed using three classification algorithms commonly used in the field of sentiment analysis, and a hyperparameter optimization method of grid search is used to discover the best parameters for each classifier. It is shown that the best performing model is the combination of Bigram feature representation, ADASYN oversampling method, and Naive Bayes classifier in the case of more positive or more negative samples. When the imbalance ratio is very large, the Naive Bayes classifier still performs the best, and the use of Borderline-SMOTE oversampling is more suitable. github: https://github.com/wenhan20201/sentiment-analysis