A Study on Predicting Ratings System through Sentiment Analysis and Clustering of Coupang Product Reviews

With the expansion of e-commerce platforms, online reviews have become increasingly important. Users often decide on purchases based on reviews and ratings. This study notes that reviews with identical ratings can differ in emotional content and hypothesizes that overall satisfaction can be inferred from these emotions. The aim is to develop a system to quantify satisfaction based on review content, especially when ratings are absent. This approach addresses situations where reviews lack ratings or users struggle to rate products despite providing feedback. We crawled product reviews from Coupang and performed sentiment analysis using a BERT model tailored for Korean. The sentiment analysis results were clustered using six different algorithms, and the outcomes were evaluated. The study explored the best methodology among 30 combinations of clustering algorithms and hyper-parameters. For evaluation, 'accuracy' and 'mean absolute error' were used as external metrics to compare with actual ratings, while the 'silhouette coefficient' was used as an internal metric to assess cluster clarity. Agglomerative Clustering showed the highest performance in external metrics but lower internal metric performance. Spectral Clustering demonstrated well-balanced and excellent performance across all metrics.