Comparative Study of Machine Learning Models for Sentiment Analysis of Amazon Product Reviews

Figure Ilustration AI

FORMOSA NEWS - Lampung - BERT Outperforms Other AI Models in Amazon Review Sentiment Analysis. Researchers from Muhammadiyah University of Lampung and Sumatra Institute of Technology have found that BERT, a transformer-based artificial intelligence model, delivers the highest accuracy in analyzing customer sentiment from Amazon product reviews. The findings were published in 2026 in the Formosa Journal of Computer and Information Science (FJCIS).  The study was conducted by Tri Noviantoro of Muhammadiyah University of Lampung and Suryaneta from Sumatra Institute of Technology. The researchers compared four widely used machine learning and deep learning models Naive Bayes, Support Vector Machine (SVM), Long Short-Term Memory (LSTM), and Bidirectional Encoder Representations from Transformers (BERT) to determine which model performs best in identifying positive and negative sentiment in online customer reviews.  The research matters because online marketplaces such as Amazon receive millions of customers reviews every day. Those reviews influence purchasing decisions, brand reputation, product development, and digital marketing strategies. Businesses increasingly rely on sentiment analysis tools to understand customer satisfaction and detect consumer concerns in real time.

AI Sentiment Analysis Becomes Essential in E-Commerce
The growth of e-commerce has transformed customer feedback into one of the most valuable business assets. Online reviews provide direct insight into customer experiences, product quality, and purchasing behavior. However, manually reading millions of reviews is impossible for most companies. Natural Language Processing (NLP), a branch of artificial intelligence focused on understanding human language, has become a key solution for processing large amounts of text data automatically. Sentiment analysis systems can classify reviews as positive, negative, or neutral, allowing businesses to identify trends and respond quickly to customer needs. Traditional machine learning methods such as Naive Bayes and SVM have long been used for text classification tasks because they are relatively simple and computationally efficient. However, modern customer reviews often contain sarcasm, contextual meaning, emotional nuance, and ambiguous language that older models struggle to interpret accurately. To address this challenge, researchers increasingly use deep learning models such as LSTM and transformer-based systems like BERT, which are designed to understand language context more effectively.

More Than One Million Amazon Reviews Analyzed
The study used the Amazon Product Reviews dataset, specifically reviews from the “All Beauty” category. After data cleaning and preprocessing, the dataset contained more than one million reviews. The researchers divided the data into training and testing groups before evaluating the four AI models.
The dataset included several types of information:

• Customer review text.
• Product ratings from one to five stars.
• Review titles.
• Posting timestamps.
• Verified purchase indicators.
• Helpful vote counts from other users.

The research team simplified the data by converting all text to lowercase, removing punctuation and irrelevant words, and transforming the reviews into machine-readable numerical representations. Reviews rated one to three stars were classified as negative sentiment, while four- and five-star reviews were categorized as positive sentiment.

BERT Achieves the Highest Accuracy
The results showed a clear performance gap between transformer-based AI and traditional machine learning models.
BERT achieved the strongest overall performance across all evaluation metrics:

• Accuracy: 92.3 percent.
• Precision: 93.22 percent.
• Recall: 94.12 percent.
• F1-score: 93.67 percent.

LSTM ranked second with an accuracy of 89.6 percent and an F1-score of 91.79 percent. Meanwhile, SVM achieved 84.5 percent accuracy, while Naive Bayes recorded 81.2 percent accuracy. According to the researchers, BERT’s advantage comes from its bidirectional transformer architecture, which allows the model to understand words based on context from both directions in a sentence. This capability helps BERT interpret emotional nuance and complex sentence structures more accurately than older machine learning approaches. The study also found that Naive Bayes and SVM generated higher false-positive rates, especially when handling long or context-dependent reviews. Although those models remain faster and less resource-intensive, they struggled to capture subtle sentiment patterns in customer feedback. LSTM performed significantly better than traditional models because it can retain long-term contextual information in sequences of text. However, it still fell slightly behind BERT in understanding complex contextual relationships between words.

Implications for Business and Digital Platforms
The findings have direct implications for e-commerce companies, digital marketplaces, and businesses that depend on customer feedback analysis.
With more accurate sentiment analysis systems, companies can:

• Monitor customer satisfaction in real time.
• Detect product issues earlier.
• Improve marketing strategies.
• Enhance product development decisions.
• Strengthen brand reputation management.

The researchers explained that BERT is highly suitable for applications requiring high precision, such as automated customer feedback analysis and online brand monitoring. However, BERT also requires substantial computational resources, including advanced GPUs and large memory capacity. In this study, the models were trained using Google Colab Pro with NVIDIA A100 GPU hardware, 32 GB RAM, and 15 GB GPU memory. For organizations with limited technological infrastructure, simpler models such as Naive Bayes and SVM may still offer practical solutions due to their lower computational costs. The researchers also noted that future studies could expand sentiment analysis into multilingual environments and integrate additional data sources such as product images and videos to create more comprehensive customer insight systems.

Author Profiles
Tri Noviantoro is a researcher from Muhammadiyah University of Lampung specializing in machine learning, natural language processing, and data analytics.
Suryaneta is an academic from Sumatra Institute of Technology with expertise in artificial intelligence, information systems, and data science.

Sources
Tri Noviantoro dan Suryaneta, Comparative Study of Machine Learning Models for Sentiment Analysis of Amazon Product Reviews. Formosa Journal of Computer and Information Science, Vol. 5 No. 1 Tahun 2026.
DOI: https://doi.org/10.55927/fjcis.v5i1.16389
URL: https://journal.formosapublisher.org/index.php/fjcis