Wide & Deep Learning is a framework that combines wide linear models and deep neural networks for recommender systems. It addresses the challenge of achieving both memorization (learning frequent item interactions) and generalization (exploring new feature combinations) in recommendation systems. The wide component uses cross-product feature transformations to memorize frequent interactions, while the deep component uses embeddings to generalize to unseen feature combinations. The two components are jointly trained to balance these two aspects. The framework was implemented and evaluated on Google Play, a mobile app store with over one billion active users and over one million apps. Online experiments showed that the Wide & Deep model significantly increased app acquisitions compared to wide-only and deep-only models. The system was productionized and open-sourced in TensorFlow. The wide component is a generalized linear model of the form y = w^T x + b, where x is a vector of features and w are the model parameters. The deep component is a feed-forward neural network that uses embeddings to represent categorical features. The two components are combined using a weighted sum of their outputs, which are then fed into a logistic loss function for joint training. The system implementation includes three stages: data generation, model training, and model serving. Data generation involves creating training data from user and app impression data. Model training involves learning embeddings for categorical features and training the wide and deep components. Model serving involves scoring app candidates and ranking them based on the model's predictions. The system was tested in an A/B testing framework for three weeks, showing a 3.9% increase in app acquisition rates compared to a wide-only model. The system also demonstrated improved serving performance, with latency reduced to 14 ms through multithreading and parallel processing. The framework was compared to previous approaches, including factorization machines and collaborative deep learning, and showed significant improvements in both memorization and generalization. The Wide & Deep model is a promising approach for recommender systems, combining the strengths of both linear and neural network models.Wide & Deep Learning is a framework that combines wide linear models and deep neural networks for recommender systems. It addresses the challenge of achieving both memorization (learning frequent item interactions) and generalization (exploring new feature combinations) in recommendation systems. The wide component uses cross-product feature transformations to memorize frequent interactions, while the deep component uses embeddings to generalize to unseen feature combinations. The two components are jointly trained to balance these two aspects. The framework was implemented and evaluated on Google Play, a mobile app store with over one billion active users and over one million apps. Online experiments showed that the Wide & Deep model significantly increased app acquisitions compared to wide-only and deep-only models. The system was productionized and open-sourced in TensorFlow. The wide component is a generalized linear model of the form y = w^T x + b, where x is a vector of features and w are the model parameters. The deep component is a feed-forward neural network that uses embeddings to represent categorical features. The two components are combined using a weighted sum of their outputs, which are then fed into a logistic loss function for joint training. The system implementation includes three stages: data generation, model training, and model serving. Data generation involves creating training data from user and app impression data. Model training involves learning embeddings for categorical features and training the wide and deep components. Model serving involves scoring app candidates and ranking them based on the model's predictions. The system was tested in an A/B testing framework for three weeks, showing a 3.9% increase in app acquisition rates compared to a wide-only model. The system also demonstrated improved serving performance, with latency reduced to 14 ms through multithreading and parallel processing. The framework was compared to previous approaches, including factorization machines and collaborative deep learning, and showed significant improvements in both memorization and generalization. The Wide & Deep model is a promising approach for recommender systems, combining the strengths of both linear and neural network models.