E-Commerce Review Quality Prediction
A machine-learning pipeline for identifying low-quality customer reviews.
Problem
Review quality cannot be inferred from text alone. The project combines behavioral, temporal, and text-derived signals to distinguish low-quality submissions.
Approach
I built the pipeline from cleaning and exploratory analysis through feature engineering, training, and held-out evaluation. Logistic Regression provided a baseline, followed by Random Forest and XGBoost comparisons.
The final evaluation reached a 0.831 ROC-AUC on the held-out test set.
Engineering decisions
- Keep training and test transformations consistent through a single feature pipeline.
- Compare a readable linear baseline with tree-based ensemble models.
- Use ROC-AUC to evaluate ranking quality across classification thresholds.
- Inspect behavioral and temporal variables alongside derived text features.
Stack
Python, pandas, NumPy, scikit-learn, XGBoost, and Matplotlib.