Skip to project

Linear algebra · 2026

Sparse SVD Recommender

From a few movie ratings to a recommendation, with each matrix operation visible along the way.

Open the demo View source

An unrated movie is not a zero-star movie

The example starts with ten users and sixteen movies. Only observed ratings contribute to each movie’s average. Known ratings become residuals around that baseline; a missing entry means no observed deviation, rather than a negative review.

Sparse input
Keep the known user–movie ratings as a coordinate list.
Movie baseline
Compute the average rating from the available observations.
Residuals
Represent how each observed rating differs from its movie’s baseline.

Follow the factorization

  1. Choose a rankExplore one to six latent components and inspect the singular values.
  2. ReconstructCombine the factors to estimate missing ratings.
  3. RecommendRank the movies the selected user has not rated.

The explanation runs in the browser

Interactive calculation

Edit the “You” ratings, change the retained rank, and follow the factors through a dot-product prediction. The static demo needs no Python server.

Python reference

A separate NumPy, SciPy and scikit-learn example expresses the same approach with a sparse residual matrix and truncated SVD.

A model you can inspect

The small table makes matrix reconstruction understandable. Its recommendations illustrate the method rather than establish recommendation quality on a real catalog or unseen users.