Analisis Sentimen Ulasan Produk Skincare Sephora Menggunakan Algoritma Bert

Abstrak

The rapid growth of the skincare industry has made consumer reviews a crucial factor in purchasing decisions. However, the massive volume of reviews on platforms like Sephora makes efficient manual analysis difficult. This study aims to develop and evaluate an automated sentiment analysis system to classify Sephora skincare product reviews into positive, negative, and neutral categories using the Bidirectional Encoder Representations from Transformers (BERT) algorithm. The research methodology applies a modified CRISP-DM framework. A dataset of 4,687 reviews was collected through crawling from the X (Twitter) platform for the period of January 2025 to January 2026. The data was cleaned into 4,683 tweets, pre-labeled using the VADER lexicon, and split with an 80:20 ratio for training and testing data. The modeling process was conducted through fine-tuning the bert-base-uncased model for 3 epochs. Experimental results demonstrate that the BERT algorithm achieved the highest performance with an accuracy of 91%, precision of 90%, recall of 89%, and an F1-score of 90%. This performance significantly outperforms baseline methods such as VADER (72% accuracy) and Naive Bayes (78% accuracy). The sentiment analysis successfully extracted consumer perception patterns; positive sentiment was dominated by service recommendation reviews, while negative sentiment focused on price complaints and product side effects. The BERT prediction results showed 2,756 tweets (58.9%) of positive sentiment, 975 tweets (20.8%) of negative sentiment, and 952 tweets (20.3%) of neutral sentiment. This pattern is consistent with the initial VADER label distribution. This research provides practical recommendations for consumers, Sephora, and manufacturers in understanding market preferences and optimizing business strategies.