In 2016 Heng-Tze Cheng and a large team at Google published Wide & Deep Learning, an architecture they put into production for app recommendations in Google Play. The paper frames recommendation as a balance between two abilities: memorization, learning that specific combinations of features tend to lead to a click, and generalization, inferring that new, unseen combinations might also work.
Their answer was to train two components jointly. The “wide” part is a linear model fed cross-product features (for example, “installed app X and is viewing app Y”), which is good at memorizing frequent, useful patterns. The “deep” part is a feed-forward neural network over learned embeddings of the same features, which generalizes to combinations the wide model has never seen. Trained together, the wide side covers the exceptions and the deep side handles the broad patterns. In a live experiment on Google Play, the combined model improved app acquisitions over using either part alone.
Wide & Deep was influential both as a model and as a design pattern, and Google released an open-source TensorFlow implementation that made the approach easy to adopt. It helped popularize deep learning for large-scale recommendation and click-through-rate prediction.
For a business reader, this paper captures a practical truth about recommenders: the best systems both remember the specific things that work and generalize to suggest things no one has tried yet.