Recommender systems often struggle to balance memorization (learning frequent, specific co-occurrences of items/features) and generalization (recommending items that haven't explicitly appeared together in the training data) [1606.07792].
This architecture has since become a standard baseline for many recommendation tasks in industry, including those described in studies on YouTube recommendations [1606.07792]. If you'd like, I can:
The paper proposes training both components simultaneously rather than separately. This allows the model to optimize for both accuracy (via the wide component) and serendipity/novelty (via the deep component) [1606.07792]. Key Results & Impact 888.470760_415140.lt.
Online experiments showed that "Wide & Deep" significantly increased app acquisitions compared to models that used either approach alone [1606.07792].
The model was heavily used for app recommendations on the Google Play Store [1606.07792]. This allows the model to optimize for both
Discuss the used in the model (e.g., user, context, item features).
A deep feed-forward neural network is used, which generalizes better to unseen feature combinations by learning low-dimensional dense embeddings for sparse features [1606.07792]. Discuss the used in the model (e
Explain the in more detail (which also uses deep learning). Find the open-source code for the Wide & Deep model.