Recommender Systems
How Netflix, YouTube, and TikTok pick what you see: collaborative filtering, matrix factorization, two-tower retrieval, ranking, and evaluation.
Before this: Machine Learning. Deep Learning helps for the later steps.
-
8 min
Step 1The math behind recommenders
The core idea of predicting ratings from similar users and items.
The Math Behind Recommender SystemsArt of the Problem
-
9 min
Step 2Content-based vs collaborative filtering
Two basic strategies and when each one works.
Content-based filtering & collaborative filtering (Building recommendation systems with TensorFlow)TensorFlow
-
12 min
Step 3Collaborative filtering in practice
User-based and item-based neighbors, with their weak spots.
Collaborative Filtering : Data Science Conceptsritvikmath
-
33 min
Step 4Matrix factorization
Learn hidden taste factors from a sparse ratings table.
How does Netflix recommend movies? Matrix FactorizationLuis Serrano Academy
-
14 min
Step 5Build a recommender in Python
Turn the ideas into working code on a movie dataset.
Building a Recommendation System in PythonSpencer Pao
-
9 min
Step 6Neural collaborative filtering
Replace the dot product with a neural network in PyTorch.
Neural Collaborative Filtering (NCF) Explanation & Implementation in PytorchSteeve Huang (steeve)
RUCAIBox/RecBole★ 4.6k -
26 min
Step 7Two-tower retrieval
How big systems narrow millions of items to a few hundred.
Building Scalable Retrieval System with Two-Tower Models | Query-Item Retrieval Recsys Model | ML AIDataTrek
tensorflow/recommenders★ 2k -
19 min
Step 8Fast nearest-neighbor search
Serve embeddings at scale with approximate search indexes.
Vector Search & Approximate Nearest Neighbors (ANN) | FAISS (HNSW & IVF)Mustafa Zaki
facebookresearch/faiss★ 41.8k -
23 min
Step 9YouTube's deep recommender paper
Walk through the classic two-stage candidate and ranking design.
[Paper Review]: Deep Neural Networks for YouTube RecommendationsAladdin Persson
-
9 min
Step 10Design a recommendation engine
See retrieval, ranking, and serving fit into one system.
Design an ML Recommendation Engine | System DesignInterview Pen
-
21 min
Step 11Ranking metrics: MRR, MAP, NDCG
Measure whether the right items land near the top.
Every Ranking Metric : MRR, MAP, NDCGritvikmath
-
12 min
Step 12Explore versus exploit with bandits
Handle new items and users while still learning from clicks.
Multi-Armed Bandit : Data Science Conceptsritvikmath
Build it
Watching isn't learning. Pick one and put it on GitHub.
- Train matrix factorization on MovieLens and report NDCG@10 against a popularity baseline.
- Build a two-tower model on MovieLens and serve its embeddings through a FAISS index.
- Write a short comparison of three recommenders on one dataset, with metrics and failure cases.