This study aims to explore the consistency between customers’ written sentiments and their numerical star ratings in Amazon reviews of beauty and personal care products. We employ three sentiment analysis models—VADER (lexicon-based), BERT (transformer-based), and LSTM (recurrent neural network)—to classify each customer review as positive, neutral, or negative. To assess the alignment between sentiment labels and star ratings, we define consistency as follows: positive sentiment corresponds to 4 or 5-star ratings, neutral sentiment to 3-star ratings, and negative sentiment to 1 or 2-star ratings. After labeling all reviews using each model, we compare the generated sentiment labels with the actual ratings provided by users. The evaluation includes visual analysis through pie charts to illustrate the distribution of predicted sentiments for each model. In addition, we compute classification metrics—precision, recall, F1-score, and support—to assess the performance of each sentiment analysis method. Confusion matrices are also provided to highlight the agreement or disagreement between predicted sentiments and star ratings. The results offer insights into how accurately sentiment analysis models capture user intent and emotion as reflected in their written feedback, and how this correlates with their explicit numerical evaluations. This research contributes to the understanding of model behavior in real-world opinion mining contexts.

Authors:
Reyhane Farshbaf Sabahi, Islamic Azad University, Iran

About the Presenter(s)
Reyhane Farshbaf Sabahi, researcher and product manager, focuses on customer behavior analysis, data-driven marketing, and machine learning. Currently working on clustering-based personalized marketing strategies.

Connect on Linkedin
https://www.linkedin.com/in/reyhane-farshbaf-sabahi-6a2552214/

See this presentation on the full schedule – Monday Schedule

Bookmark