Matrix Factorization (SVD)

Matrix Factorization SVD decomposes the user-item rating matrix into latent user and item factor vectors. At inference time, it predicts ratings by computing the dot product of the user and item latent vectors, then ranks items by predicted rating.

When to use:

• Collaborative filtering on explicit rating data (star ratings, scores)
• Personalized recommendations based on user-item interaction history
• When latent factor representations of users and items are useful for downstream tasks

Input: User-item interaction data (user ID, item ID, optionally ratings) Output: Ranked list of recommended items per user with predicted scores

Model Settings (set during training, used at inference)

Number of Factors (default: 100) Dimensionality of the latent user and item vectors. More factors capture richer preferences at the cost of training time and overfitting risk.

Number of Epochs (default: 20) Training iterations over the rating matrix.

Learning Rate (default: 0.005) SGD step size for factor updates.

Regularization (default: 0.02) L2 regularization on latent factors. Prevents overfitting on sparse rating matrices.

Use Bias (default: true) Whether to include user and item bias terms to capture global rating tendencies.

Inference Settings

No dedicated inference-time settings. Recommendations are ranked by predicted rating score.