Probabilistic Topic Modeling dan Perbandingan Algoritma Analisis Sentimen dengan Pendekatan Hybrid pada Ulasan Wardah Glasting Liquid Lip

Abstract

Wardah menjadi merek lokal dengan penjualan lipstik terbanyak di Shopee pada Juni 2024. Salah satu produknya yaitu Wardah Glasting Liquid Lip. Penelitian ini bertujuan mengidentifikasi topik utama pada ulasan Wardah Glasting Liquid Lip di Shopee menggunakan metode Latent Dirichlet Allocation dan membandingkan performa tiga algoritma analisis sentimen yaitu Support Vector Machine, CatBoost Classifier, dan Long Short-Term Memory dengan pendekatan hybrid. Perbandingan didasarkan pada accuracy, precision, recall, dan F1-score. Pelabelan sentimen dilakukan secara manual dan word embedding menggunakan IndoBERT. Proses topic modeling menghasilkan tiga topik utama yaitu tekstur, pengiriman, dan warna. Topik dengan sentimen positif terbanyak yaitu warna dan sentimen negatif terbanyak yaitu tekstur. Model Long Short-Term Memory menjadi model terbaik dengan accuracy sebesar 73,2%. Akan tetapi, ketiga model masih mengalami kesulitan dalam mengklasifikasikan sentimen netral dengan rendahnya nilai F1-score pada kelas tersebut yang hanya berada di rentang 14-29%.

Wardah became the local brand with the highest lipstick sales on Shopee in June 2024. One of its products is the Wardah Glasting Liquid Lip. This study aims to identify the main topics in reviews of the Wardah Glasting Liquid Lip on Shopee using the Latent Dirichlet Allocation method and to compare the performance of three sentiment analysis algorithms: Support Vector Machine, CatBoost Classifier, and Long Short-Term Memory using a hybrid approach. The comparison is based on accuracy, precision, recall, and F1-score. Sentiment labeling was performed manually, while word embedding utilized IndoBERT. The topic modeling process revealed three predominant topics: texture, shipping, and color. The topic with the most positive sentiment was color, while the most negative sentiment was texture. The Long Short-Term Memory model performed best, achieving an accuracy of 73,2%. However, all three models struggled to classify neutral sentiment, as indicated by the low F1-score in that class which ranged from just 14% to 29%