Skip to main content

SVD (Matrix Factorization)

Description​

The SVD (Singular Value Decomposition) policy implements matrix factorization, often referring to iterative algorithms inspired by SVD used for collaborative filtering (like Funk SVD or SVD++). It decomposes the user-item interaction matrix into latent factors and potentially incorporates user/item biases (mean-centering tendencies) to improve prediction accuracy. It's typically trained using optimization techniques like Stochastic Gradient Descent (SGD).

Policy Type: svd Supports: embedding_policy, scoring_policy

Hyperparameter tuning​

  • factors: Number of latent factors in the matrix factorization.
  • num_epochs: Number of complete passes through the training dataset.

V1 API

policy_configs:
  embedding_policy: # Can also be used under scoring
    policy_type: svd
    n_factors: 100        # Number of latent factors 
    n_epochs: 20          # Number of training iterations (epochs) for SGD
    biased: true          # Whether to include user/item bias terms 
    lr_all: 0.005         # Learning rate for SGD
    reg_all: 0.02         # Regularization strength for SG

Usage​

Use this model when:​

  • You have explicit feedback data (ratings, reviews with scores)
  • You want to model user and item biases (mean-centering tendencies)
  • You prefer SGD-based optimization over ALS
  • You need a straightforward matrix factorization approach
  • You're working with medium-sized datasets

Choose a different model when:​

  • You have only implicit feedback (ALS is typically better for this)
  • You need to leverage item content features (use Two-Tower or BeeFormer)
  • You want the most scalable solution (ELSA may be better)
  • You need to model sequential patterns (use sequential models)

Use cases​

  • Product recommendations with review scores
  • Restaurant or service recommendations with ratings
  • Any domain with explicit user feedback

References​