Build a Recommender from Scratch

What you'll be able to do

✓Build a recommender from scratch — similarity, collaborative filtering, and matrix factorization over a real ratings dataset
✓Represent sparse interaction data honestly, so a rating that was never given is never treated as a rating of zero
✓Split ratings by time and measure how much a random split flatters your results
✓Establish baselines before modelling, so you always know what beating nothing looks like
✓Measure recommendations with Recall@K, NDCG@K, MAP and catalog coverage instead of rating error alone
✓Train matrix factorization by gradient descent over observed entries, the way the Netflix Prize was won
✓Show why the best rating predictor is not the best recommender, using your own measurements

Build the system behind every "recommended for you" row you have ever seen — and, more importantly, the evaluation harness that tells you honestly whether it works, which is the part most tutorials skip and most production teams get wrong.

⌁ Amazon's item-to-item recommendations, the Netflix Prize solutions, and Spotify's early collaborative filtering all run on the techniques you build here; every modern recommender is measured against exactly these baselines.

Start this internship

Create an account to unlock the 13 sections, the workbench, and AskThili.

Begin