Deep learning for recommender systems

Description

Recommendation systems are increasingly important in our digital world, and many organizations are looking to machine learning to create recommendations that leverage their unique datasets to provide better results. Most applications have used techniques best described as wide learning: logistic regression, minhash, and weighted alternated least squares (WALS). As part of this trend, many leading digital companies are using deep learning to create recommendation models, simplifying feature engineering and resulting in improved performance.

Tim Seears and Karthik Bharadwaj Thirumalai explain how to apply deep learning to improve consumer recommendations by training neural nets to learn categories of interest using embeddings. They then demonstrate how to extend this with WALS matrix factorization to achieve wide and deep learning—a process which is now used in production for the Google Play Store.

Topics include:

• Fundational concepts for wide and deep models
• Deep learning models using neural nets with embeddings
• Measuring model performance
• Visualizing model behavior
• Wide learning models using TensorFlow MLToolkit’s WALS matrix factorization
• Deep and wide learning combining both techniques
• Ideas for further applications of deep and wide learning