Sentiment analysis of Amazon product reviews

Abstract

The rapid development of online shopping sites has pushed people's shopping to a new way. Online shopping not only provides convenience to people but also "suggestions." Moreover, there are always many reviews from previous consumers on shopping websites, helping people know more about the product and make decisions. This paper represents the sentiment analysis of Amazon reviews using three models: Random Forest, Naive Bayes, and SVM. These models are trained with token counts, and term frequency-inverse document frequency (TF-IDF) features to make better comparisons. Classification performances are evaluated by precision, recall, and F-1 scores, and exploration is implemented into the dataset providing information about Amazon reviews. The results show that Random Forest and SVM models perform well on positive-labeled data but provide suboptimal results on negative-labeled and neutral-labeled data. Overall, Naive Bayes has the best performance for all three classifications. However, classifications might be biased during the analysis. Thus, more improvements are expected in future research about this topic to obtain more accurate and ideal results, and more machine learning models are supposed to be implemented.