Retail Rocket Re-commender System

  • Tech Stack: Python, SQL, AWS Lambda, XGBoost, Neural Networks
  • Link: Project Link

Built an end-to-end recommendation system using advanced machine learning models trained on over 1 million real-world retail transactions. The project aimed to predict product preferences and drive smarter, faster recommendations using a hybrid of XGBoost and Neural Networks, enabling personalized user experiences in real time. The system was deployed via AWS Lambda to ensure serverless, scalable model execution and integrated into a simulated e-commerce flow to test latency, accuracy, and user response rates. Evaluation centered on improving recommendation relevance and system responsiveness.

Key features of the project include:

  • XGBoost + Neural Net Ensemble combined gradient boosting with deep learning to capture both structured patterns and non-linear behaviors.
  • Generated real-time suggestions for cart additions, similar items, and cross-category bundles.
  • Used SQL for preprocessing, feature engineering, and session-based behavioral metrics.
  • Reduced infrastructure overhead and cut response times by 30% on average.
  • R² Score of 0.69, delivered strong model performance for a multi-class recommendation setup.

Key Insights

  • User session length, product views per session, and purchase frequency strongly impacted model accuracy.
  • AWS Lambda reduced traditional model latency from ~1.5s to ~1.0s, enabling smoother real-time UX.
  • Neural Nets enhanced prediction quality for long-tail products, while XGBoost dominated high-volume categories.
  • Training time was optimized using categorical encoding, feature pruning, and batch updates to accelerate convergence.