<![CDATA[In the digital era, product reviews are an important source of information for consumers and businesses because they influence purchasing decisions and marketing strategies. However, the distribution of sentiment in product reviews is often unbalanced, with positive reviews dominating and negative reviews being limited. This condition poses a challenge in developing text classification models, especially for Indonesian which has a complex morphological structure and very rich vocabulary variations. This study adapts the Contrastive Learning method for the classification of unbalanced Indonesian language product reviews and tests the effectiveness of text augmentation techniques in improving representation, especially for minority classes with limited data. Data were obtained through web scraping from Indonesian e-commerce platforms, totaling around 10,000 reviews with a composition of 52% positive, 30% negative, and 18% neutral. The data was processed and expanded using augmentation techniques to significantly increase the variety and amount of training data. The LSTM model trained on the original data and the augmented data, showing an increase in validation accuracy from around 73% to almost 100% in the 30th epoch, with a final accuracy reaching 92% and an F1-Score of 90%. These results confirm that the incorporation of data augmentation is crucial to address imbalance, thereby improving the robustness and reliability of the model in product review sentiment classification]]>
