Amazon Reviews NLP Analysis

Project Context

Developed during the “Automatic Data Analysis” course (SAE S5), this project leverages a real-world dataset (Amazon Reviews 2023 - Software) to perform classification and regression tasks.

The goal is to predict a user’s rating (1 to 5 stars) based solely on the text of their review.

Technical Pipeline

We built a comprehensive pipeline to transform raw text into actionable predictions.

1. Data Preparation & NLP

• Cleaning: Natural Language Processing using spaCy (tokenization, stop-word removal).
• Vectorization: Converting text comments into numerical data.
• Rebalancing: Applied Stratified Undersampling (via imbalanced-learn) to address the dataset’s heavy bias towards positive reviews.
• Normalization: Used StandardScaler to scale features appropriately.

2. Classification Models

We trained and compared several algorithms to classify reviews (positive vs. negative, or multi-class):

• K-Nearest Neighbors (KNN)
• Decision Tree
• Random Forest (Selected as the best performing model)
• Fine-tuning: Hyperparameter search using GridSearchCV integrated within Scikit-learn Pipelines.
• Comparison: Bar charts comparing Macro F1-Scores to identify the optimal text/model configuration.

3. Optimization & Regression

• Linear Regression: Predicted exact scores (RMSE) based on textual indicators.
• Residual Analysis: Used scatter plots and histograms to analyze prediction errors (Actual vs. Predicted ratings).
• Performance: Evaluated via RMSE on test and validation sets to measure numerical model precision.

Results Visualization

To evaluate model performance, we generated:

• F1-Score: To measure the balance between precision and recall for each rating category.
• Confusion Matrices to analyze classification errors.
• Precision-Recall Curves to verify robustness against class imbalance.
• Residual Histograms to validate error distribution in regression.
• RMSE Comparison across different preprocessing methods (RAW vs. LEMMA).