Course contentsShow
← Machine Learning and Deep Learning
Lesson 2367 of 3,53851. Recommender SystemsFree lesson

Wide & Deep Networks for Recommendations

Combine memorization and generalization using wide linear models with deep neural components jointly trained.

Wide & Deep Networks for Recommendations

What You'll Learn

You'll learn how to combine memorization of specific patterns with generalization across user-item interactions by jointly training wide linear models with deep neural networks.

The Core Idea

Imagine a restaurant recommender that needs to both remember that "users who ordered item A always order item B" (memorization) and generalize that "users who like spicy food tend to enjoy Asian cuisine" (generalization). Wide & Deep networks achieve both simultaneously.

The wide component is a linear model that learns direct feature interactions—essentially memorizing co-occurrence patterns through cross-product transformations. Think of it as capturing "if feature X AND feature Y, then recommend Z."

The deep component is a multi-layer neural network (like the NCF architectures you've learned) that learns low-dimensional embeddings and discovers hidden patterns through non-linear transformations.

How They Work Together

Instead of choosing between matrix factorization's generalization or feature-based memorization, you train both components jointly with a combined loss function:

final_prediction = sigmoid(wide_output + deep_output)

The wide path gets sparse, crossed features (like "user_app=Netflix AND genre=Action"), while the deep path gets user and item embeddings that flow through dense layers. Both paths contribute to the final prediction, and gradients flow back through both during training.

This architecture excels when you have both:

  • Rich categorical features (app installs, browsing history) → wide side
  • User/item identifiers needing embeddings → deep side

Key Takeaway: Wide & Deep networks combine the memorization power of linear models for specific feature patterns with the generalization capability of deep networks for latent factors, training them jointly to capture both explicit rules and hidden relationships in recommendation systems.