Score prediction from e-commerce product reviews

This study aims to perform sentiment analysis and star rating estimation on Turkish product reviews using machine learning and natural language processing techniques using user reviews on Trendyol, one of Turkey's leading e-commerce platforms. Product reviews made by users on e-commerce platforms and the star ratings given to these reviews play an important role in consumers' learning about products. However, inconsistencies between reviews and star ratings make the average ratings of products misleading and can have negative consequences for both consumers and sellers. Within the scope of this study, it is aimed to develop a more reliable rating estimation system by detecting and correcting these inconsistencies between Turkish product reviews and star ratings. In the study, first a dataset containing user reviews and star ratings was created, and then data preprocessing steps were performed. Five different methods were used to detect inconsistencies and analyzes were conducted to determine the most effective method. For feature extraction, texts were converted to numerical representation using Term Frequency-Inverse Document Frequency (TF-IDF) and Word2Vec algorithms. Then, machine learning algorithms such as Random Forest, Logistic Regression, Support Vector Machine (SVM) and Gradient Boosting were trained and tested with hyperparameter optimizations. The analyses were performed on twoclass (positive, negative), three-class (positive, negative, neutral) and five-class classification models. The results obtained showed that correcting the inconsistencies increased the model accuracy and more reliable star rating estimation could be made. This study is an important step in improving the user experience on e-commerce platforms, minimizing misleading content and developing more accurate decision support systems.