Personalizing user experiences with electronic content based on user representations learned from application usage data

Patent metadata

Field	Value
Publication number	US-10614381-B2
Application number	US-201615381637-A
Country	US
Kind code	B2
Filing date	Dec 16, 2016
Priority date	Dec 16, 2016
Publication date	Apr 7, 2020
Grant date	Apr 7, 2020

Abstract

This disclosure involves personalizing user experiences with electronic content based on application usage data. For example, a user representation model that facilitates content recommendations is iteratively trained with action histories from a content manipulation application. Each iteration involves selecting, from an action history for a particular user, an action sequence including a target action. An initial output is computed in each iteration by applying a probability function to the selected action sequence and a user representation vector for the particular user. The user representation vector is adjusted to maximize an output that is generated by applying the probability function to the action sequence and the user representation vector. This iterative training process generates a user representation model, which includes a set of adjusted user representation vectors, that facilitates content recommendations corresponding to users' usage pattern in the content manipulation application.

First claim

The invention claimed is:

1. A method for learning, from application usage data, user representations that are usable for personalizing user experiences with electronic content, the method comprising:
accessing, by a computing system and from a memory device, training action histories for users of a content manipulation application, wherein each training action history comprises a respective sequence of user actions performed by a respective user with the content manipulation application;
accessing, by the computing system, a user representation model, wherein the user representation model includes a probability function and user representation vectors indicating respective software usage patterns for the users, wherein the user representation vectors have respective initial values;
training, by the computing system, the user representation model via iterations, wherein training the user representation model comprises, for each iteration of at least some of the iterations:
selecting, from a respective training action history for a respective user, respective training data comprising (i) a respective target action, (ii) a respective set of previous actions performed before the respective target action, and (iii) a respective set of subsequent actions performed after the respective target action, wherein each of the respective target action, the respective set of previous actions and the respective set of subsequent actions is represented by an action vector encoding a corresponding action,
computing a respective initial output by applying the probability function to action vectors of the respective set of previous actions and the respective set of subsequent actions, and a respective user representation vector for the respective user,
adjusting at least the respective user representation vector for the respective user and one or more of the action vectors of the respective set of previous actions and the respective set of subsequent actions to increase a value of an output of the probability function based on the respective initial output being less than a respective maximized output of the probability function; and

outputting a set of adjusted user representation vectors from the trained user representation model, wherein the adjusted user representation vectors improves the user representation model for predicting target actions in the content manipulation application from sequences of actions in the content manipulation application.

2. The method of claim 1 , wherein the user representation model further comprises (i) a weight vector, (ii) an action representation matrix having action vectors representing available actions in the content manipulation application and including the action vectors of the respective target action, the respective set of previous actions, and the action vectors of the respective set of subsequent actions, and (iii) a user representation matrix having the user representation vectors, wherein each sequence of user actions in the training action histories comprises a respective subset of the action vectors,
wherein the initial outputs are computed, in the iterations, by applying the probability function to the weight vector, the action representation matrix, and the user representation matrix in addition to the action vectors of the respective set of previous actions and the respective set of subsequent actions, and the respective user representation vector for the respective user, and
wherein training the user representation model comprises adjusting values of the weight vector, adjusting values of the action vectors used to represent the available actions in the action representation matrix, and adjusting values of the user representation vectors used to represent usage patterns in the user representation matrix, wherein the weight vector, the action representation matrix, and the user representation matrix are adjusted based on at least some initial outputs of the probability function being less than corresponding maximized outputs of the probability function for at least some of the training action histories.

3. The method of claim 2 , wherein, for each iteration, the probability function computes a respective probability of the respective target action using a transfer function that combines at least the respective user representation vector, and a respective subset of the action vectors representing the respective set of previous actions and the respective set of subsequent actions.

4. The method of claim 3 , wherein the probability function computes each probability by applying a softmax function to a transfer function output computed with the transfer function, wherein the probability function computes the initial outputs and the maximized outputs from a logarithmic function that is respectively applied to the probabilities.

5. The method of claim 1 , further comprising:
learning, with the computing system, a projection function that transforms the adjusted user representation vectors into recommendation user representation vectors, respectively, that are usable with a recommendation model for computing recommendation values indicating user preferences for content items accessible via an online content service that is separate from the content manipulation application;
accessing, by the computing system, an input user representation vector that is computed from an input action history in the content manipulation application;
transforming, with the computing system, the input user representation vector into a transformed user representation vector by applying the projection function to the input user representation vector; and
generating, with the computing system, a content recommendation in the online content service by applying the recommendation model to the transformed user representation vector.

6. The method of claim 5 , further comprising computing the recommendation user representation vectors, wherein computing each recommendation user representation vector comprises:
identifying, based on a respective training viewing history from the online content service, a respective set of content items that has been previously accessed, via the online content service, by a respective user associated with the respective training viewing history;
identifying a user-item loss function that relates the respective set of content items to (i) a respective initial recommendation user representation vector of the respective user and (ii) a respective set of item representation vectors representing the respective set of content items;
minimizing the user-item loss function by modifying the respective initial recommendation user representation vector into the recommendation user representation vector and adjusting dimension values in the respective set of item representation vectors, wherein minimizing the user-item loss function optimizes the recommendation model,
wherein the optimized recommendation model computes, from the recommendation user representation vector and the respective set of item representation vectors with the adjusted dimension values, training recommendation values that are proportional to preferences of the respective user for the respective set of content items.

7. The method of claim 6 , wherein learning the projection function comprises minimizing a projection loss function that relates the recommendation user representation vector to a training vector that is computed by applying the projection function to a corresponding one of the user representation vectors from the trained user representation model.

8. The method of claim 6 , wherein generating the content recommendation comprises computing, with the optimized recommendation model, output recommendation values from the transformed user representatio

Assignees

• Adobe Inc

Inventors

• Hoffman Matthew
• Yang Longqi
• Jin Hailin
• Fang Chen

Classifications

• G06Q30/0255

  based on user history · CPC title

• H04W12/06

  Authentication · CPC title

• G06Q30/0277

  Online advertisement · CPC title

• G06N20/00 Primary

  Machine learning · CPC title

• G06N7/005

  Physics · mapped topic