Predicting Customer Review Ratings in Fashion E-Commerce: A Machine Learning Approach with Enhanced Feature Engineering

The fashion industry is notoriously dynamic, marked by rapid trend shifts and volatile consumer behaviour. This project shifts focus from traditional sales forecasting to predicting customer review ratings, leveraging textual data and associated metadata from e-commerce platforms. The primary objective was to develop a stable and highly accurate machine learning framework that predicts binned product ratings based on comprehensive feature engineering and advanced Natural Language Processing (NLP) techniques. Key feature engineering steps included creating sentiment scores, developing aspect-based features, and strategically categorizing demographic data (age into generational groups). Utilizing the Women's Clothing E-Commerce Reviews dataset, the research rigorously compares a traditional Random Forest Classifier against a deep learning Bidirectional Long Short-Term Memory (LSTM) network. The results conclusively demonstrate that the LSTM model significantly outperforms the strong baseline, achieving superior stability (Std Dev of 0.0019 and, critically, a vastly improved ability to identify the minority class of negative or neutral reviews, F1= 0.7134 vs. 0.47. This validates the need for deep sequential learning in this domain, providing a nuanced, valuable, and stable predictive tool for proactive retail strategic planning and quality control.