1 Google deep & wide app recommender system
Figure : Wide & Deep model structure.
1 Input Features
·Deep
Including continuousand enumerated features. Enumerated features are changed toembedding, randomly initialized and convergence after deep&wide model’s training, as MLP’s inputs.
For example:
Continuous features such as app installed age arediscretized by quantile to [0,1].
·Wide
History items cross product current item to classification. E.g. History items are {A, B, C}, current item is {D}.(A and D) is 1 if user satisfy A and D simultaneously, otherwise 0.
·Compare of Deep and Wide
Deep can find hidden features through relative low dimension features, but has more complex computing.
Wide with elaborately designing high dimension features can reach good result.. It’s simple, scalable, and interpretable.
2 Experiment
+3.9%relative to the control group (statistically significant).