Wide & Deep Learning for Recommender Systems

This paper presents Wide & Deep learning for recommender systems, which jointly trains a wide linear model and a deep neural network. The wide component uses cross-product feature transformations to memorize feature interactions in an effective and interpretable way, though this typically demands substantial feature engineering. The deep component learns low-dimensional dense embeddings for sparse features, letting it generalize to previously unseen feature combinations with less feature engineering, but such models can over-generalize and recommend less relevant items when user-item interactions are sparse and high-rank. Combining both aims to capture the benefits of memorization and generalization together.

The authors productionized and evaluated the system on Google Play, a commercial mobile app store with over one billion active users and more than one million apps. In online experiments, Wide & Deep significantly increased app acquisitions compared with wide-only and deep-only models. They also open-sourced their implementation in TensorFlow, demonstrating that jointly training wide and deep components can combine memorization and generalization in a production recommender at scale.