Wide & Deep: app recommendations in Google Play
On 24 June 2016 sixteen authors at Google posted the Wide & Deep preprint: a wide linear model on cross-product features and a deep network with embeddings, trained jointly. The model was put into production and tested in Google Play; in an online experiment on 1 per cent of users it raised app acquisitions from the store's main landing page by 3.9 per cent against the previous model, a wide-only logistic regression.
Why it matters
The work showed how an industrial system combines memorising frequent feature combinations with generalising to rare ones, and gave an online measurement in a product with more than a billion active users. The implementation was open-sourced in TensorFlow, so others could take up the architecture.
Preprint v1: Google Play has 'over one billion active users and over one million apps'; training on over 500 billion examples; the deep part is three ReLU layers over a vector of about 1,200 dimensions; serving in about 10 ms per request. A second 1 per cent group got the deep part alone, and Wide & Deep gained 1 per cent over it. Both online figures are the company's own measurement; there is no independent confirmation.