Method and system for domain-adaptive content suggestion for an online concierge system
An online concierge system uses a domain-adaptive suggestion module to score products that may be presented to a user as suggestions in response to a user's search query. The domain-adaptive suggestion module receives data that is relevant to scoring products as suggestions in response to a search query. The domain-adaptive suggestion module uses one or more domain-neutral representation models to generate a domain-neutral representation of the received data. The domain-neutral representation is a featurized representation of the received data that can be used by machine-learning models in the search domain or the suggestion domain. The domain-adaptive suggestion module then scores products by applying one or more machine-learning models to domain-neutral representations generated based on those products. By using domain-neutral representations, the domain-adaptive suggestion module can be trained based on training examples from a similar prediction task in a different domain.
1. A suggestion engine stored on one or more non-transitory computer-readable storage media, wherein the suggestion engine is manufactured by a process comprising:
accessing training data that comprises a plurality of examples, wherein the plurality of examples comprises:
a set of suggestion examples, wherein each suggestion example comprises a feature set describing an instance where a product is presented to a user as a suggestion and a label describing whether the user interacted with the suggestion; and
a set of search examples, wherein each search example comprises a feature set describing an instance where a product is presented to a user as a search result and a label describing whether the user interacted with the search result;
accessing a domain-adaptive suggestion module that comprises:
a domain-neutral representation module comprising one or more neural networks that generate a domain-neutral representation from user data, search query data, and product data;
a suggestion scoring module comprising one or more neural networks that generate a suggestion score based on the domain-neutral representation that is output from the domain-neutral representation module, wherein a suggestion score represents an affinity of a product to be presented as a suggestion;
a search scoring module comprising one or more neural networks that generate a search score based on the domain-neutral representation output from the domain-neutral representation module, wherein a search score represents an affinity of a product to be presented as a search result; and
a domain-neutrality scoring module comprising one or more neural networks that output a domain-neutrality score from a domain-neutral representation, the domain-neutrality score indicating an ability of the domain-neutrality scoring module to predict whether the domain-neutral representation was generated from a suggestion example or a search example;
training the suggestion engine by repeatedly:
selecting an example from the plurality of examples;
generating a domain-neutral representation by the domain-adaptive suggestion module based on the selected example;
responsive to selecting a suggestion example, updating parameters of the one or more neural networks of the domain-neutral representation module and parameters of the one or more neural networks of the suggestion scoring module based on the domain-neutral representation;
responsive to selecting a search example, updating the parameters of the one or more neural networks of the domain-neutral representation module and parameters of the one or more neural networks of the search scoring module based on the domain-neutral representation; and
updating the parameters of the one or more neural networks of the domain-neutral representation module based on a domain-neutrality score generated by the domain-neutrality scoring module for the domain-neutral representation;
stopping the training when one or more criteria are met; and
storing the parameters of the one or more neural networks of the domain-neutral representation module and the parameters of the one or more neural networks of the suggestion scoring module on the one or more computer-readable media as parameters of the suggestion engine.
2. The suggestion engine of claim 1 , wherein each suggestion example in the set of suggestion examples comprises:
a feature set describing a user associated with the suggestion example;
a feature set describing a search query associated with the suggestion example; and
a feature set describing a suggestion associated with the suggestion example.
3. The suggestion engine of claim 1 , wherein each search example in the set of search examples comprises:
a feature set describing a user associated with the search example;
a feature set describing a search query associated with the search example; and
a feature set describing a set of products presented as search results associated with the search example.
4. The suggestion engine of claim 1 , wherein domain-neutral representations generated by the domain-neutral representation module comprise one or more of a feature vector or an embedding.
5. The suggestion engine of claim 1 , wherein a suggestion score generated by the suggestion scoring module represents a likelihood of a user to interact with a suggestion based on a product associated with the suggestion score.
6. The suggestion engine of claim 1 , wherein a search score generated by the search scoring module represents a likelihood of a user to interact with a search result based on a product associated with the search score.
7. The suggestion engine of claim 1 , wherein a suggestion score represents an affinity of a product to be presented to a user in response to a search query from the user.
8. The suggestion engine of claim 1 , wherein the domain-neutrality scoring module comprises a discriminator network.
9. The suggestion engine of claim 1 , wherein the domain-neutrality scoring module generates domain-neutrality scores based on a difference between a distribution representing the set of search examples and a distribution representing the set of suggestion examples.
10. The suggestion engine of claim 1 , wherein the process for manufacturing the suggestion engine further comprises storing the parameters of the one or more neural networks of the search scoring module.
11. A method comprising:
accessing training data that comprises a plurality of examples, wherein the plurality of examples comprises:
a set of suggestion examples, wherein each suggestion example comprises a feature set describing an instance where a product is presented to a user as a suggestion and a label describing whether the user interacted with the suggestion; and
a set of search examples, wherein each search example comprises a feature set describing an instance where a product is presented to a user as a search result and a label describing whether the user interacted with the search result;
accessing a domain-adaptive suggestion module that comprises:
a domain-neutral representation module comprising one or more neural networks that generate a domain-neutral representation from user data, search query data, and product data;
a suggestion scoring module comprising one or more neural networks that generate a suggestion score based on the domain-neutral representation that is output from the domain-neutral representation module, wherein a suggestion score represents an affinity of a product to be presented as a suggestion;
a search scoring module comprising one or more neural networks that generate a search score based on the domain-neutral representation output from the domain-neutral representation module, wherein a search score represents an affinity of a product to be presented as a search result; and
a domain-neutrality scoring module comprising one or more neural networks that output a domain-neutrality score from a domain-neutral representation, the domain-neutrality score indicating an ability of the domain-neutrality scoring module to predict whether the domain-neutral representation was generated from a suggestion example or a search example;
training a suggestion engine by repeatedly:
selecting an example from the plurality of examples;
generating a domain-neutral representation by the domain-adaptive module based on the selected example;
responsive to selecting a suggestion example, updating parameters of the one or more neural networks of the domain-neutral representation module and parameters of the one or more neural networks of the suggestion scoring module based on the domain-neutral representation;
responsive to selecting a search example, updating the parameters of the one or more neural networks of the domain-neutral representation module and parameters of the one or more neural networks of the search scoring module based on the domain-neutral representation; and
updating the parameters of the one or more neural networks of the domain-neutral representation module based on a domain-neutrality score generated by the domain-neutrality scoring module for the domain-neutral representation;
stopping the training when one or more criteria are met; and
storing the parameters of the one or more neural networks of the domain-neutral representation module and the parameters of the one or more neural networks of the suggestion scoring module on the one or more computer-readable media as parameters of the suggestion engine.
12. The method of claim 11 , wherein each suggestion example in the set of suggestion examples comprises:
a feature set describing a user associated with the suggestion example;
a feature set describing a search query associated with the suggestion example; and
a feature set describing a suggestion associated with the suggestion example.
13. The method of claim 11 , wherein each search example in the set of search examples comprises:
a feature set describing a user associated with the search example;
a feature set describing a search query associated with the search example; and
a feature set describing a set of products presented as search results associated with the search example.
14. The method of claim 11 , wherein domain-neutral representations generated by the domain-neutral representation module comprise one or more of a feature vector or an embedding.
15. The method of claim 11 , wherein a suggestion score generated by the suggestion scoring module represents a likelihood of a user to interact with a suggestion based on a product associated with the suggestion score.
16. The method of claim 11 , wherein a search score generated by the search scoring module represents a likelihood of a user to interact with a search result based on a product associated with the search score.
17. The method of claim 11 , wherein a suggestion score represents an affinity of a product to be presented to a user in response to a search query from the user.
18. The method of claim 11 , wherein the domain-neutrality scoring module comprises a discriminator network.
19. The method of claim 11 , wherein the domain-neutrality scoring module generates domain-neutrality scores based on a difference between a distribution representing the set of search examples and a distribution representing the set of suggestion examples.
20. The method of claim 11 , further comprising storing the parameters of the one or more neural networks of the search scoring module.