<![CDATA[This paper presents a systematic sentiment analysis framework for Indonesian-language e-commerce reviews, designed for scalable extraction of insights for digital product development. The system applies a Support Vector Machine (SVM) classifier with a Radial Basis Function (RBF) kernel and TF-IDF bigram feature extraction to the PRDECT-ID dataset comprising 5,400 product reviews from Tokopedia across 29 categories. To ensure reliable classification under realistic class conditions, the proposed pipeline integrates multi-stage text preprocessing (case folding, slang normalization, stopword removal, and Sastrawi-based stemming) and stratified 80:20 train-test splitting without artificial resampling. The experimental evaluation confirmed a test-set accuracy of 92.04%, a weighted F1-score of 0.92, and an AUC-ROC of 0.9741. These results validate the efficacy of the proposed SVM-TF-IDF architecture for reliable, interpretable sentiment classification. Furthermore, category-level negative sentiment profiling identifies Computers and Laptops, Automotive, and Toys and Hobbies as priority intervention domains (negative rate ?60%), while keyword-level TF-IDF analysis reveals critical user concerns regarding delivery services, product specification mismatches, and quality disappointment, providing tangible guidance for product development teams. Future work should explore transformer-based architectures (BERT, IndoBERT) for contextual sentiment capture, investigate cross-marketplace generalizability, and address real-time deployment scalability.]]>
