Knowledge Discovery of Application Review Using Word Embedding's Comparison with CNN-LSTM Model on Sentiment Analysis

The rapid growth of e-commerce in Indonesia reflects the country's economic acceleration and provides a variety of benefits. One of them is TikTok Shop, which has become a popular platform for digital economic activities that offers opportunities for businesses and consumers. To maximize these opportunities, companies must effectively understand and manage user sentiment to improve services and user experience. Accurate sentiment analysis plays a crucial role in achieving this goal. This research focuses on comparing the effectiveness of various word embedding methods in enhancing sentiment analysisof TikTok Shop user reviews. The study applies a combined CNN-LSTM approach, comparing classic pre-trained embeddings like GloVe with custom-trained embeddings using Word2Vec and FastText, which are trained on domain-specific data. The objective of this research is to identify the most effective embedding method for capturing user sentiment and providing more accurate insights. The findings show that the custom-trained technique, Word2Vec, delivered the best performance, achieving the highest accuracy, precision, F1-score, and AUC-ROC values. FastText ranked second with nearly the same performance, differing only slightly. In contrast, GloVe pre-trained ability to capture user sentiment is not as strong as that of custom-trained Word2Vec and FastText, although it remains quite effective overall.