Efficient Modeling and Prediction of E-commerce Recommendation Systems: A Case Study on Amazon Product Reviews

Project Overview

This project explores the predictive potential of e-commerce review data by analyzing Amazon beauty product reviews (2004–2018). It employs machine learning to predict product ratings and builds a recommendation system to provide insights into customer preferences and product popularity.

Key Features:

• Rating Prediction: Binary classification of reviews (5-star vs. 1-4 stars).
• Recommendation System: Collaborative filtering based on reviewer-product engagement.
• Feature Engineering: TF-IDF vectorization, sentiment scoring, and datetime transformations.
• Dimensionality Reduction: Principal Component Analysis (PCA) to improve efficiency and prevent overfitting.

Team Members

• Yulong Dong
• Daniel Sitompul
• Khanh Thai

Project Deliverables

• Research Question: Can customer reviews predict product ratings?
• Models Used: Logistic Regression, Random Forest, and Stacking Models.
• Performance Metrics: ROC Curve, AUC, Confusion Matrix, and RMSE.
• Video Presentation: Watch here

Dataset

The dataset includes 5.2k reviews from Amazon beauty products, featuring:

• Review Text
• Summary
• Product Style
• Reviewer Information
• Ratings and Votes

Data Source: UCSD's open-source Amazon Review dataset.

Methodology

1. Exploratory Data Analysis: Identifying missing data and reviewing feature distributions.
2. Feature Engineering:
   • Textual Data: Sentiment analysis and TF-IDF vectorization.
   • Temporal Data: Transforming Unix timestamps to datetime features.
3. Model Training:
   • Logistic Regression as baseline.
   • Random Forest for non-linear classification.
   • Dimensionality reduction using PCA.
4. Recommendation System: Built using collaborative filtering and Singular Value Decomposition (SVD).

Key Results

• Model Performance:
  • Random Forest outperformed Logistic Regression, achieving an AUC of 0.99.
  • PCA reduced model complexity by 85% while preserving accuracy.
• Recommendation System:
  • Provided highly correlated product suggestions based on user-product engagement.

Contact

For questions or feedback, contact:

• Daniel Sitompul: daniel_sitompul@berkeley.edu

---

Let me know if you need further assistance!