Overview
Matrix factorization learns latent row and column representations whose product reconstructs observed entries. It is often trained by minimizing reconstruction error on known values while regularizing the factors.
CoreSource checked
Matrix Factorization
Approximates a data matrix as a product of lower-dimensional latent factor matrices.
Quick Facts
Compact metadata for implementation and discovery.
- Category
- Machine Learning
- Difficulty
- Intermediate
- Family
- Latent Factor Models
- Tier
- Core
- Status
- Verified
- Review
- Source Checked
- Sources
- 1
- Formulas
- 0
- Relationships
- 4
- Code example
- Ready
- Visualization
- Ready
Core Concepts
recommendation modelinglatent factor modelsfactorizationmachine learning
Quick facts
Complexity
Runtime and memory behavior at a glance.
Best
O(k)
Average
O(k^2)
Worst
O(k^3)
Space
O(k)
When to Use
- Use when interactions are sparse but low-rank structure is plausible.
- Use when latent factors are useful for prediction or compression.
When Not to Use
- Avoid when missing entries are not missing at random and bias is unmodeled.
- Avoid when cold-start rows or columns have no observations.
What It Does
- Represents rows and columns with latent vectors
- Predicts entries from vector dot products
- Optimizes factors against observed data
Use Cases
- Collaborative filtering and recommender systems
- Low-rank approximation
- Imputing missing matrix entries under a latent-factor model
Inputs
- Sparse observed interaction matrix R (users × items)
- Latent rank k, learning rate η, and regularization λ
Outputs
- Low-rank factor matrices U (m × k) and V (n × k)
- Completed dense score predictions R_hat = U V^T
Pseudocode
- 1.Initialize row and column factor matrices
- 2.For each observed entry
- 3.Predict the entry from the factor dot product
- 4.Compute reconstruction error
- 5.Update both factor vectors with regularization
- 6.Validate predictions on held-out observed entries
Canonical Example
- Input
- ratings for users U1,U2 and items I1,I2 with one missing rating
- Output
- latent user/item vectors and a predicted missing rating
- Explanation
- The model fits vectors so known dot products match observed ratings and then predicts the missing entry.
Implementation Notes
- Split observed entries into train and validation sets.
- Initialize factors with small random values or a stable decomposition.
- Track reconstruction loss and regularization separately.
Common Pitfalls
- Evaluating on training entries hides overfitting.
- Unregularized factors can memorize sparse observations.
- Scale and bias terms should be handled consistently.
Interactive Mechanism Walkthrough
Iterative Convergence & State TrajectoryShow decomposing an observed matrix into low-rank latent factors on a small recommendation modeling example.
Step 1 of 6: Initialize row and column factor matrices
machine-learning-mechanismNotes
Matrix factorization is a modeling assumption; validation should check whether low-rank structure actually helps.
Sources (1)
What each reference supports for this entry.
Matrix Factorization Techniques for Recommender Systems
PaperReference for low-rank matrix factorization with user and item latent factors in recommendation.
Open sourceExplore Next
Good next jumps from the relationship graph, ranked for useful discovery.
Often paired withStochastic Gradient Descent
Optimization · Core
Large matrix-factorization models are often trained with stochastic gradient updates.
Compare with thisRelationships
View full relationship pageConceptual neighborPrincipal Component Analysis
Both use low-dimensional linear structure, but matrix factorization often handles sparse observed entries.
Often paired withStochastic Gradient Descent
Large matrix-factorization models are often trained with stochastic gradient updates.
Related Algorithms
Tags
Machine LearningLinear Algebra