Content presentation based on a multi-task neural network

Patent metadata

Field	Value
Publication number	US-10803377-B2
Application number	US-201615053448-A
Country	US
Kind code	B2
Filing date	Feb 25, 2016
Priority date	Feb 25, 2016
Publication date	Oct 13, 2020
Grant date	Oct 13, 2020

Abstract

Techniques for predictively selecting a content presentation in a client-server computing environment are described. In an example, a content management system detects an interaction of a client with a server and accesses client features. Responses of the client to potential content presentations are predicted based on a multi-task neural network. The client features are mapped to input nodes and the potential content presentations are associated with tasks mapped to output nodes of the multi-task neural network. The tasks specify usages of the potential content presentations in response to the interaction with the server. In an example, the content management system selects the content presentation from the potential content presentations based on the predicted responses. For instance, the content presentation is selected based on having the highest likelihood. The content management system provides the content presentation to the client based on the task corresponding to the content presentation.

First claim

The invention claimed is:

1. A computer-implemented method comprising:
detecting, by a content management system, an interaction of a client with a server, the content management system configured for customizing content presentations for the client based on the interaction of the client with the server;
accessing, by the content management system, features associated with the client based on the interaction of the client with the server;
predicting, by the content management system, responses of the client to potential content presentations based on a multi-task neural network, wherein:
the features of the client correspond to an input layer of the multi-task neural network,
the potential content presentations correspond to tasks of an output layer of the multi-task neural network,
the tasks specify usages of the potential content presentations, and
the multi-task neural network is trained based on historical data associated with the potential content presentations through backpropagation, wherein a backpropagation gradient from one of the tasks having a non-existence label is set to zero;

selecting, by the content management system, a content presentation from the potential content presentations based on the responses;
providing, by the content management system to the client for presentation, the content presentation based on a task of the multi-task neural network corresponding to the content presentation;
adding, by the content management system, a new task to the multi-task neural network, the new task corresponding to a new potential content presentation; and
incrementally retraining, by the content management system, the multi-task neural network based on a set of responses of a set of clients to the new potential content presentation rather than the historical data associated with the potential content presentations.

2. The computer-implemented method of claim 1 , wherein selecting the content presentation is further based on a rule, wherein the rule specifies a selection based on a predicted likelihood of a response to the content presentation.

3. The computer-implemented method of claim 1 , wherein selecting the content presentation is further based on a rule, wherein the content management system stores the rule based on input received from a computing device of an entity that is associated with content of the content presentation, and further comprising:
tracking metrics about selections of the potential content presentations; and
providing the metrics to the computing device of the entity, wherein the providing enables the entity to manage usage of the potential content presentations via the computing device.

4. The computer-implemented method of claim 1 , wherein selecting the content presentation is further based on a comparison of the responses to the potential content presentations.

5. The computer-implemented method of claim 1 , wherein a hidden layer of the multi-task neural network is configured to transform the features associated with the client to a set of shared features across the tasks.

6. The computer-implemented method of claim 1 , a hidden layer of the multi-task neural network is configured to apply a non-linear transformation to the features associated with the client.

7. The computer-implemented method of claim 1 , wherein the multi-task neural network comprises a hierarchy of hidden layers between the input layer and the output layer.

8. The computer-implemented method of claim 1 , wherein the multi-task neural network is configured to apply a non-linear transformation to the features associated with the client, wherein the non-linear transformation is selected from potential non-linear transformations based on a cross-validation that utilizes a training dataset.

9. The computer-implemented method of claim 1 , further comprising:
tracking, by the content management system, historical data that comprises historical responses of clients to the potential content presentations; and
training, by the content management system, the multi-task neural network based on the historical data.

10. The computer-implemented method of claim 1 , further comprising:
tracking, by the content management system, the responses of the set of clients to the new potential content presentation for retraining the multi-task neural network.

11. A system comprising:
a processor;
a memory communicatively coupled to the processor and bearing instructions that, upon execution by the processor, cause the system to at least:
detect a request from a computing device of a user for a web page;
access features associated with the user based on the request;
predict user responses to candidate variations of the web page based on a multi-task neural network, wherein
the features associated with the user correspond to an input layer of the multi-task neural network,
the candidate variations of the web page correspond to tasks of an output layer of the multi-task neural network,
the tasks specify usage of the candidate variations of the web page, and
the multi-task neural network is trained based on historical data associated with the candidate variations of the web page through backpropagation, wherein a backpropagation gradient from one of the tasks having a non-existence label is set to zero;

select a variation of the web page from the candidate variations of the web page based on the user responses;
provide the variation of the web page for presentation at the computing device based on a task of the multi-task neural network corresponding to the variation;
add a new task to the multi-task neural network, the new task corresponding to a new variation of the web page; and
incrementally retrain the multi-task neural network based on a set of responses of a set of users to the new variation of the web page rather than the historical data associated with the candidate variations of the web page.

12. The system of claim 11 , wherein a user response to the variation comprises a user conversion that is based on a presentation of the variation of the web page at the computing device.

13. The system of claim 11 , wherein predicting the user responses comprises determining likelihoods of user conversions.

14. The system of claim 11 , wherein selecting the variation of the web page comprises determining that a user response to the variation corresponds to the highest likelihood of a user conversion relative to remaining responses.

15. The system of claim 11 , wherein the features associated with the user comprise information about the user, information about the computing device of the user, and information about the web page requested from the computing device of the user.

16. The system of claim 11 , wherein the variation of the web page comprises a particular advertisement to be displayed on the web page or a particular location within the web page to display an advertisement.

17. A non-transitory computer-readable storage medium storing instructions that, when executed on a system, configure the system to perform operations comprising:
detecting an interaction of a client with a server;
accessing features associated with the client based on the interaction;
predicting responses of the client to potential presentations of content at the client based on a multi-task neural network, wherein
the features associated with the client correspond to an input layer of the multi-task neural network,
the potential presentations of content correspond to tasks of an output layer of the multi-task neural network,
the tasks specify usage of the potential present

Assignees

• Adobe Inc

Inventors

• Roychowdhury Anirban
• Bui Trung
• Kucera John
• Bui Hung
• Jin Hailin

Classifications

• H04L67/02 Primary

  based on web technology, e.g. hypertext transfer protocol [HTTP] · CPC title

• H04W4/23

  for mobile advertising · CPC title

• G06N3/082

  modifying the architecture, e.g. adding, deleting or silencing nodes or connections · CPC title

• G06N3/084

  Backpropagation, e.g. using gradient descent · CPC title

• G06Q30/0277

  Online advertisement · CPC title