Factorized models

1. A computer-implemented method performed by data processing apparatus, the method comprising:

training a machine learned model to predict an outcome, the training being on a set of query-resource pairs, wherein each query-resource pair corresponds to a respective query and a respective resource observed as having been identified as being responsive to the respective query, the training generating a plurality of tokens that each describe at least a pair of features derived from the query-resource pair and a constituent weight for the outcome for which the machine learned model is trained, and wherein a likelihood of the outcome for a given query and given resource of a query-resource pair is based on the constituent weights of the tokens that correspond to the given query and given resource;

for each of the queries of the query-resource pairs, generating a respective query feature vector having a cardinality of elements, each element corresponding to a query feature value of a feature to be learned;

for each of the resources of the query-resource pairs, generating a respective resource feature vector having the cardinality of elements, each element corresponding to a resource feature value of a feature to be learned; and

training a factorization model to learn a query feature value for each element of each query feature vector and a resource feature value for each element of each resource feature vector such that the factorization model output, based on the respective feature vectors of a given query and given resource, is predictive of outcome for which the machine learned model is trained.

2. The computer-implemented method of claim 1 , wherein the factorization model predicts the outcome based on inner products of the query feature vectors and resource feature vectors.

3. The computer-implemented method of claim 1 , wherein the query features and the resource features are emergent from the training of the factorization model.

4. The computer-implemented method of claim 1 , wherein generating a respective query feature vector having cardinality of elements comprises generating a respective query feature vector for only queries that are not navigational queries.

5. The computer-implemented method of claim 1 , wherein generating a respective query feature vector having cardinality of elements comprises generating a respective query feature vector for only queries that are categorical queries.

6. The computer-implemented method of claim 1 , further comprising:

determining a set of queries that is a proper subset of queries received at a search system;

for each of the queries in the set of queries:

determining a set of resources determined to be responsive to the query;

selecting a query feature vector for the query;

selecting a resource feature vector for each resource in the set of resources;

generating a respective query-resource score based on an inner product of the query feature vector and the resource feature vector of the query and resource;

storing the respective query-resource score for the query-resource pair in a search system cache; and

for each query received by the search system that matches a query in the set of queries, retrieving the query-resource scores for the query for a search operation.

7. The computer-implemented method of claim 6 , wherein determining a set of queries that are a proper subset of queries comprises determining a set of queries that are most popular relative to other queries received at the search system.

8. The computer-implemented method of claim 1 , further comprising:

receiving, for a query from a user device, data identifying a set of resources that are determined to be responsive to the query and ranked according to a first order;

determining for each resource, from the factorization model, a likelihood of the outcome based on an inner product of the query feature vector for the query learned from the factorization model and the resource feature vector for the resource learned from the factorization model;

adjusting the ranking of the resources based, in part, on the determined likelihood of the outcome so that the set of resources are ranked according to a second order; and

providing, to the user device, data identifying the resources according to the second order.

9. The computer-implemented method of claim 8 , wherein the resources are applications, and the outcome is an installing of an application on the user device.

10. The computer-implemented method of claim 8 , wherein the resources are audio files, and the outcome is a purchase of the audio file by a user.

11. The computer-implemented method of claim 8 , wherein:

determining for each resource, from the factorization model, a likelihood of the outcome based on an inner product of the query feature vector for the query learned from the factorization model and the resource feature vector for the resource learned from the factorization model comprises determining the likelihood for only query and resource pairs for which the learned model was not trained; and

further comprising determining, from the learned model, the likelihood of the outcome for each query and resource pair for which the learned model was trained.

12. A computer storage medium encoded with a computer program, the program comprising instructions that when executed by a data processing apparatus cause the data processing apparatus to perform operations comprising:

training a machine learned model to predict an outcome, the training being on a set of query-resource pairs, wherein each query-resource pair corresponds to a respective query and a respective resource observed as having been identified as being responsive to the respective query, the training generating a plurality of tokens that each describe at least a pair of features derived from the query-resource pair and a constituent weight for the outcome for which the machine learned model is trained, and wherein a likelihood of the outcome for a given query and given resource of a query-resource pair is based on the constituent weights of the tokens that correspond to the given query and given resource;

for each of the queries of the query-resource pairs, generating a respective query feature vector having a cardinality of elements, each element corresponding to a query feature value of a feature to be learned;

for each of the resources of the query-resource pairs, generating a respective resource feature vector having the cardinality of elements, each element corresponding to a resource feature value of a feature to be learned; and

training a factorization model to learn a query feature value for each element of each query feature vector and a resource feature value for each element of each resource feature vector such that the factorization model output, based on the respective feature vectors of a given query and given resource, is predictive of outcome for which the machine learned model is trained.

13. A system comprising:

a data processing apparatus; and

a computer storage medium encoded with a computer program, the program comprising instructions that when executed by the data processing apparatus cause the data processing apparatus to perform operations comprising:

accessing a machine learned model trained to predict an outcome, the training being on a set of query-resource pairs, wherein each query-resource pair corresponds to a respective query and a respective resource observed as having been identified as being responsive to the respective query, the training generating a plurality of tokens that each describe at least a pair of features derived from the query-resource pair and a constituent weight for the outcome for which the machine learned model is trained, and wherein a likelihood of the outcome for a given query and given resource of a query-resource pair is based on the constituent weights of the tokens that correspond to the given query and given resource;

for each of the queries of the query-resource pairs, generating a respective query feature vector having a cardinality of elements, each element corresponding to a query feature value of a feature to be learned;

for each of the resources of the query-resource pairs, generating a respective resource feature vector having the cardinality of elements, each element corresponding to a resource feature value of a feature to be learned; and

training a factorization model to learn a query feature value for each element of each query feature vector and a resource feature value for each element of each resource feature vector such that the factorization model output, based on the respective feature vectors of a given query and given resource, is predictive of outcome for which the machine learned model is trained.

14. The system of claim 13 , wherein the factorization model predicts the outcome based on inner products of the query feature vectors and resource feature vectors.

15. The system of claim 13 , wherein the query features and the resource features are emergent from the training of the factorization model.

16. The system of claim 13 , wherein generating a respective query feature vector having cardinality of elements comprises generating a respective query feature vector for only queries that are not navigational queries.

17. The system of claim 13 , wherein generating a respective query feature vector having cardinality of elements comprises generating a respective query feature vector for only queries that are categorical queries.

18. The system of claim 13 , the operations further comprising:

determining a set of queries that a is proper subset of queries received at a search system;

for each of the queries in the set of queries:

determining a set of resources determined to be responsive to the query;

selecting a query feature vector for the query;

selecting a resource feature vector for each resource in the set of resources;

generating a respective query-resource score based on an inner product of the query feature vector and the resource feature vector of the query and resource;

storing the respective query-resource score for the query-resource pair in a search system cache; and

for each query received by the search system that matches a query in the set of queries, retrieving the query-resource scores for the query for a search operation.

19. The system of claim 17 , wherein the resources are applications, and the outcome is an installing of an application on the user device.