🎥 Netflix Recommender

Our team was tasked to built a Netflix Recommendation System using restricted Boltzmann machine (RBM). As this was a competition we had to pit our prediction scores with other teams as well as a baseline Linear Regression model.

• Context + Inspiration: Recommendation systems are becoming more popular from Youtube to Spotify and Online Shopping. We are provided with a dataset that contains information about user’s opinions of movies in the form of a rating. The task lies in finding what users might rate for movies they have not yet watched

• The plain RBM model severely underperformed against the Linear Regression model. We subsequently improved the accuracy by adding a few extensions. The below are the list of extensions I contributed to:

  • Random Searching
  • Momentum
  • Biases
  • L1 Regularisation
  • Adaptive Learning Rate
  • Mini-Batches

Challenges: It was a bottomless pit, a lot of the resources I found did not prove helpful as it did not translate to an improvement in prediction. This is mainly due to poor implementation. If there were more time, I will try using a Bayesian optimisation approach as well as consider rounding off values to the nearest integer to simulate a user’s pick

• Technologies used: Vanilla Python