Personalized Recommender System
This research-driven project delivers a personalized recommender system designed to simulate real-world discovery platforms like Spotify or Netflix. It predicts user preferences by analyzing historical interaction data using a combination of statistical modeling and causal reasoning.
University Research Project
September 2023
December 2023
Recommender Systems, Causal Inference
Python, NumPy, Pandas, Matplotlib
The system combines collaborative filtering, matrix factorization, and causal graphical models to learn from large-scale user-item interactions. Posterior distributions and learning curves are visualized to explain model behavior, while RMSE and MAE are used to evaluate recommendation accuracy.
- Processed over 1 million user-artist interaction records, reducing sparsity and normalizing engagement data.
- Built memory-based and model-based collaborative filtering models from scratch using matrix factorization with SVD.
- Achieved consistent reduction in RMSE and strong generalization on validation data.
- Used DAGs to explore how user preferences causally influence item popularity.
- Produced interpretable visualizations of latent dimensions and model learning dynamics.
Project Overview
This project simulates the dynamics of personalized recommendation engines found in streaming platforms. The challenge was to deliver high-precision suggestions from sparse user behavior data while maintaining interpretability and generalization.
The model pipeline includes data cleaning, transformation, exploratory analysis, CF-based recommendation generation, latent feature mapping through matrix factorization, and model validation using RMSE, MAE, and coverage metrics.
A unique aspect of this project was the use of causal inference to uncover how user interactions drive item visibility and engagement, an approach that adds an additional layer of intelligence to traditional collaborative filtering systems.
The result is a robust, explainable, and scalable system capable of generating dynamic personalized content recommendations and serving as a strong base for commercial-scale recommender systems.