Course contentsShow
← Machine Learning and Deep Learning
Lesson 2363 of 3,53851. Recommender SystemsFree lesson

From Matrix Factorization to Neural Networks

Understand the transition from classical collaborative filtering to neural architectures and their advantages.

From Matrix Factorization to Neural Networks

What you'll learn: How recommender systems evolved from linear matrix factorization to flexible neural architectures that capture complex user-item interactions.

The Transition Point

Classical collaborative filtering using matrix factorization (like SVD and ALS) represents users and items as low-dimensional vectors whose dot product approximates ratings. While powerful, this linear interaction is limiting—it assumes the relationship between user and item features is simply additive.

Think of matrix factorization as fitting a flat plane through your data points. It works well for many cases, but what if the relationship between "user loves action movies" and "item is an action movie" isn't just multiplication? What if certain combinations matter more, or interact in non-linear ways?

Neural Networks as Function Approximators

Neural networks can learn arbitrary non-linear mappings between user-item pairs and predictions. Instead of constraining yourself to rating ≈ user_vector · item_vector, you can model rating ≈ f(user, item) where f is a multi-layer neural network.

The architecture typically:

  • Takes user and item IDs as input
  • Embeds them into dense vectors (similar to latent factors in matrix factorization)
  • Passes concatenated or combined embeddings through hidden layers with non-linear activations
  • Outputs a predicted rating or ranking score

Key Advantages

Neural approaches offer:

  • Non-linear interactions: Capture complex patterns matrix factorization misses
  • Feature integration: Easily incorporate side information (user demographics, item metadata, temporal context)
  • Scalability: Modern frameworks handle large-scale training efficiently
  • Transfer learning: Pre-trained embeddings from one domain can bootstrap another

Key Takeaway: Neural networks extend matrix factorization from linear dot products to learnable non-linear functions, enabling richer representations and easier integration of auxiliary features at the cost of increased complexity.