Netflix Recommendation System Case Study

Dataset

• Netflix Prize Dataset:
  • User Ratings: Anonymous movie ratings from users between 1999 and 2005.
  • Data Files: combined_data_1.txt, combined_data_2.txt, combined_data_3.txt, and combined_data_4.txt.
  • Features: Movie ID, User ID, Rating, Date.

Problem Statement

• Objective: Predict the rating a user would give to a movie they haven’t rated yet.
• Metrics: Minimize RMSE (Root Mean Squared Error) and MAPE (Mean Absolute Percentage Error).

Models Implemented

1. Collaborative Filtering:

   • Explored user-user and item-item similarity using cosine similarity.
   • Applied Matrix Factorization using Truncated SVD for dimensionality reduction and faster similarity computation.

2. Content-Based Filtering:

   • Integrated movie metadata like titles and release dates for content-based recommendations.
   • Used TF-IDF for text-based similarity on movie descriptions.

3. Hybrid Approach:

   • Combined collaborative and content-based filtering to improve recommendation accuracy.

Results

• Truncated SVD: Reduced dimensionality of user-movie matrix, speeding up computations.
• RMSE Improvement: Models achieved significant reduction in RMSE over baseline Cinematch performance.

Conclusion

The hybrid model using collaborative filtering with SVD-based dimensionality reduction, combined with content-based filtering, provided the best accuracy and efficiency in predicting movie ratings.