Module 8 - Recommender Systems

The Production Reality

Every major consumer product runs a recommender system. Netflix drives 80% of watch time through recommendations. Amazon's item-to-item CF generates 35% of revenue. Spotify's Discover Weekly keeps users from churning. Recommendation is not a nice-to-have feature - it is often the core product.

This module covers the full arc: from the foundational math of neighborhood methods and matrix factorization, through neural collaborative filtering and two-tower retrieval, to the learning-to-rank systems that score and order candidates before they reach the user. You will learn how these systems are built in production, where they break, and how to explain every design decision in an interview.

Module Map

Lesson Guide

#	Lesson	Core Concept
01	Collaborative Filtering	User-based and item-based neighborhood methods; cosine similarity; Pearson correlation; scaling via item-item precomputation
02	Content-Based Filtering	Item profiles from features; TF-IDF; user preference vectors; hybrid approaches
03	Matrix Factorization	SVD, ALS, BPR; latent factor models; the Netflix Prize breakthrough
04	Neural Collaborative Filtering	Replacing dot products with MLPs; embedding layers; GMF + MLP fusion
05	Two-Tower Models	Separate user and item encoders; approximate nearest neighbor retrieval; YouTube DNN architecture
06	Learning to Rank	Pointwise, pairwise, and listwise objectives; LambdaMART; NDCG optimization
07	Cold Start Problem	New user and new item challenges; content bootstrapping; exploration-exploitation

Key Concepts at a Glance

The recommendation problem. Given a set of users $U$ and items $I$, and a sparse observed interaction matrix $R \in \mathbb{R}^{|U| \times |I|}$, predict the unobserved entries (explicit feedback) or rank unobserved items by probability of interaction (implicit feedback). The challenge: $R$ is typically 99%+ missing. You must generalize from a tiny fraction of observed behavior to a complete ranking over thousands or millions of items.

Three families of approaches. (1) Collaborative filtering uses only interaction data - users who behaved similarly in the past will behave similarly in the future. No item metadata required. (2) Content-based filtering uses item and user features - text, images, categories - to compute similarity without requiring any interaction history. (3) Hybrid and neural methods combine both signals in learned embedding spaces, enabling end-to-end optimization over the full recommendation objective.

Production at scale. Real recommendation systems are almost never a single model. They are multi-stage pipelines: a retrieval stage that selects hundreds of candidates from millions of items (two-tower, ANN search), a ranking stage that scores and orders those candidates (gradient boosted trees, deep ranking models), and a re-ranking stage that applies business rules, diversity constraints, and freshness signals. Understanding this pipeline architecture is as important as knowing any individual algorithm.

:::note Prerequisites This module assumes familiarity with linear algebra (matrix operations, dot products, SVD), basic probability (expectation, conditional probability), and Python with NumPy. Lessons 04 and 05 require comfort with PyTorch and neural network training. :::