Recommendation system with sparse feature encoding
A recommendation system includes a recommendation model for generating a recommendation score for a user with respect to an item. The model is configured to receive a set of dense features, describing numerical information, and a set of sparse features, representing a subset of items from a relatively large group of items. To represent the subset of items in the sparse features, each item (or a symbol thereof) is processed by an encoder to represent each item with a plurality of positions in a sparse binary representation of the subset of items. The sparse binary representation is then processed by a model that determines a vector representation of the sparse category features used in the prediction in conjunction with the dense features.
1. A method for recommending an item to a user, the method comprising:
identifying a set of dense features and a set of sparse category features associated with recommending the item to the user, the set of sparse category features describing a subset of symbols belonging to an alphabet of category symbols;
determining, by a dense model, a dense vector from the set of dense features;
determining a sparse binary representation of the set of sparse features with an encoder applied to each of the symbols of the set of sparse category features;
determining, by a category feature model, a category vector representation from the sparse binary representation; and
inputting the dense vector and the category vector representation into a prediction model, the prediction model outputting a recommendation score of the item for the user; and
providing a recommendation of the item for the user based on the recommendation score,
wherein the dense model, category feature model, and prediction model are trained neural networks.
2. The method of claim 1 , wherein each symbol in the alphabet of category symbols represents a different item in a category.
3. The method of claim 1 , wherein the sparse binary representation has a plurality of positions and wherein the encoder applied to a symbol designates a plurality of indices; and wherein determining the sparse binary representation comprises setting positions of the sparse binary representation to a value of one if an index of the position is included in the plurality of indices for the encoder applied to any of the symbols of the set of sparse category features.
4. The method of claim 1 , wherein the encoder includes a plurality of hash functions.
5. The method of claim 1 , further comprising:
receiving a request from the device for a content recommendation for the user,
wherein the recommendation is provided to the device in response to the request.
6. The method of claim 5 , wherein the set of sparse features includes items associated with the user in a category of the alphabet of category symbols.
7. The method of claim 1 , wherein the dense features and the sparse category features are included in a set of training data, wherein the method further comprises further training the category feature model based on the recommendation score and the set of training data.
8. A system comprising:
a processor; and
a non-transitory computer-readable storage medium storing instructions executable by the processor to perform operations for recommending an item to a user, the operations comprising:
identifying a set of dense features and a set of sparse category features associated with recommending the item to the user, the set of sparse category features describing a subset of symbols belonging to an alphabet of category symbols,
determining, by a dense model, a dense vector from the set of dense features,
determining a sparse binary representation of the set of sparse features with an encoder applied to each of the symbols of the set of sparse category features,
determining, by a category feature model, a category vector representation from the sparse binary representation;
inputting the dense vector and the category vector representation into a prediction model, the prediction model outputting a recommendation score of the item for the user, and
providing a recommendation of the item for the user to a device based on the recommendation score,
wherein the dense model, category feature model, and prediction model are trained neural networks.
9. The system of claim 8 , wherein each symbol in the alphabet of category symbols represents a different item in a category.
10. The system of claim 8 , wherein the sparse binary representation has a plurality of positions and wherein the encoder applied to a symbol designates a plurality of indices; and wherein determining the sparse binary representation comprises setting positions of the sparse binary representation to a value of one if an index of the position is included in the plurality of indices for the encoder applied to any of the symbols of the set of sparse category features.
11. The system of claim 8 , wherein the encoder includes a plurality of hash functions.
12. The system of claim 8 , wherein the operations further comprise:
receiving a request from the device for a content recommendation for a user,
wherein the recommendation is provided to the device in response to the request.
13. The system of claim 12 , wherein the set of sparse features includes items associated with the user in a category of the alphabet of category symbols.
14. The system of claim 8 , wherein the dense features and the sparse category features are included in a set of training data and the operations further comprise further training the category feature model based on the recommendation score and in the set of training data.
15. A non-transitory computer-readable storage medium storing instructions executable by a processor to perform operations for recommending an item to a user, the operations comprising:
identifying a set of dense features and a set of sparse category features associated with recommending the item to the user, the set of sparse category features describing a subset of symbols belonging to an alphabet of category symbols;
determining, by a dense model, a dense vector from the set of dense features;
determining a sparse binary representation of the set of sparse features with an encoder applied to each of the symbols of the set of sparse category features;
determining, by a category feature model, a category vector representation from the sparse binary representation;
inputting the dense vector and the category vector representation into a prediction model, the prediction model outputting a recommendation score of the item for the user; and
providing a recommendation of the item for the user to a device based on the recommendation score,
wherein the dense model, category feature model, and prediction model are trained neural networks.
16. The non-transitory computer-readable storage medium of claim 15 , wherein each symbol in the alphabet of category symbols represents a different item in a category.
17. The non-transitory computer-readable storage medium of claim 15 , wherein the sparse binary representation has a plurality of positions and wherein the encoder applied to a symbol designates a plurality of indices; and wherein determining the sparse binary representation comprises setting positions of the sparse binary representation to a value of one if an index of the position is included in the plurality of indices for the encoder applied to any of the symbols of the set of sparse category features.
18. The non-transitory computer-readable storage medium of claim 15 , wherein the encoder includes a plurality of hash functions.
19. The non-transitory computer-readable storage medium of claim 15 , wherein the operations further comprise:
receiving a request from the device for a content recommendation for a user,
wherein the recommendation is provided to the device in response to the request.
20. The non-transitory computer-readable medium of claim 19 , wherein the set of sparse features includes items associated with the user in a category of the alphabet of category symbols.