Optimasi Multinomial Naive Bayes untuk Klasifikasi Rating Tokopedia

Abstract

Sentiment analysis of e-commerce reviews faces challenges due to the use of informal language (slang) and class imbalance, which can reduce the performance of classification models. This study aims to optimize the Multinomial Naive Bayes algorithm for classifying Tokopedia reviews into five ordinal rating classes. The proposed approach integrates lexicon-based hybrid slang normalization, TF-IDF n-gram feature representation, data balancing using the Synthetic Minority Over-sampling Technique (SMOTE), and hyperparameter optimization. The model was evaluated using Stratified 10-Fold Cross Validation on 37,146 valid reviews. The experimental results show that the proposed approach improved the average Macro F1-score by 102.42%, increasing from 0.1704 to 0.3450, while achieving an accuracy of 58%. These findings indicate that the combination of slang normalization, TF-IDF n-gram, SMOTE, and hyperparameter optimization effectively enhances the model's ability to recognize minority classes in imbalanced e-commerce review datasets.