Sentiment analysis plays a critical role in understanding customer opinions, boosting market confidence, and guiding decision-making for businesses and governments. This study presents a comprehensive comparative analysis of sentiment classification architectures for large-scale Persian e-commerce reviews. A dataset consisting of 93,868 labeled reviews was used to evaluate traditional machine learning models (Support Vector Machine and Random Forest), deep learning models with static Word2Vec embeddings (Convolutional Neural Network and Bi-directional Long Short-Term Memory), and a transformer-based contextual model (ParsBERT). Performance was assessed using accuracy, macro precision, macro recall, and macro F1 score across varying training data proportions to examine scalability and data efficiency. Results demonstrate a clear generational progression in performance. Traditional machine learning models exhibit stable but early-saturating behavior. Static embedding-based deep learning models significantly improve scalability and balanced class performance. ParsBERT achieves the highest overall accuracy (86.47%) and macro F1 score while also demonstrating superior performance in low-data regimes, reflecting the benefits of contextual transformer-based representations. The findings highlight the importance of representation learning for sentiment analysis in morphologically rich languages such as Persian and provide practical guidance for deploying scalable sentiment monitoring systems in e-commerce environments.