Session 8: Matrix Factorization & AutoRec

⚠️ Please note we are currently reworking the lecture and still preparing the course materials. You will find this message wherever the course materials are being updated / not finalized. ⚠️

Part 1 — SVD (from scratch)¶

• Motivation: raw vectors are sparse and noisy — discover latent structure

• Matrix factorization: decompose R ≈ U · Sigma · V^T

• Truncated SVD: low-rank approximation

• SGD with regularization (from scratch)

• Key insight: rows of U and V are user/item embeddings in the same latent space

Part 2 — AutoRec (with RecBole)¶

• AutoRec: the neural version of SVD — compress sparse rating vector through an autoencoder, reconstruct to fill missing entries

• Train loss only on observed ratings (mask out missing)

• RecBole introduction: train AutoRec on MovieLens with a few lines of config

• Compare SVD vs. AutoRec metrics — students see the linear-to-neural transition

Practice: SVD from scratch on MovieLens, then train AutoRec in RecBole, compare reconstruction quality and recommendation metrics