<![CDATA[Background: The rapid growth of e-commerce in Indonesia has generated massive volumes of user-generated reviews. A critical mismatch exists between numerical star ratings and the actual sentiment expressed in review texts, creating unreliable signals for platform management and highlighting the need for text-based sentiment analysis. Objective: This study aims to analyze user review sentiments towards two leading e-commerce platforms in Indonesia using Machine Learning and Deep Learning approaches. Methods: The analysis process was conducted through the CRISP-DM stages, including data cleaning, labeling, model training, and performance evaluation. Two types of labeling were used, namely Rating-Based and VADER-Based, to compare the accuracy levels of sentiment classification. Four models were applied: Random Forest, Support Vector Machine (SVM), Logistic Regression, and IndoBERT. VADER labeling was adapted for Indonesian through preprocessing with an Indonesian-English translation layer. Results: Based on the evaluation results, the IndoBERT model showed the best performance on e-commerce X, with an accuracy of 0.96 using VADER-based labeling. Meanwhile, for e-commerce Y, Random Forest achieved an accuracy of 0.81 using VADER labeling. These results indicate that IndoBERT's Transformer architecture with contextual embeddings enabled superior understanding of Indonesian semantic nuances. Random Forest's advantage on e-commerce Y (627 samples per label) reflects a lower overfitting risk compared to deep learning models on small datasets. Conclusion: This study demonstrates the effectiveness of combining IndoBERT and VADER in Indonesian sentiment analysis and can serve as a reference for the e-commerce industry to improve service quality and customer satisfaction strategies.]]>
