PREDICTIVE ANALYSIS OF TARGET BEHAVIORS UTILIZING RNN-BASED USER EMBEDDINGS

• US 20200134300A1
• Filed: 12/20/2019
• Published: 04/30/2020
• Est. Priority Date: 11/16/2017
• Status: Active Grant

• Current Assignee: Adobe Inc.
• Original Assignee: Adobe Inc.
• Inventors: KIM, SUNGCHUL, JAIN, DEEPALI, GUPTA, DEEPALI, KOH, EUNYEE, KVETON, BRANISLAV, SHEORAN, NIKHIL, SINHA, ATANU, BUI, HUNG HAI, CHEN, CHARLES LI
• Granted Patent: US 10,783,361 B2
• Time in Patent Office: Days
• CPC Class Codes: G06F 16/954 Navigation, e.g. using categorised browsing
• CPC Class Codes: G06F 18/2411 based on the proximity to a decision surface, e.g. support vector machines
• CPC Class Codes: G06N 3/042 Knowledge-based neural networks; Logical representations of neural networks
• CPC Class Codes: G06N 3/044 Recurrent networks, e.g. Hopfield networks
• CPC Class Codes: G06N 3/049 Temporal neural networks, e.g. delay elements, oscillating neurons or pulsed inputs
• CPC Class Codes: G06N 3/08 Learning methods
• CPC Class Codes: G06N 3/088 Non-supervised learning, e.g. competitive learning
• CPC Class Codes: G06V 40/20 Movements or behaviour, e.g. gesture recognition recognition of facial expressions G06V40/16

Abstract

Systems and methods provide for generating predictive models that are useful in predicting next-user-actions. User-specific navigation sequences are obtained, the navigation sequences representing temporally-related series of actions performed by users during navigation sessions. To each navigation sequence, a Recurrent Neural Network (RNN) is applied to encode the navigation sequences into user embeddings that reflect time-based, sequential navigation patterns for the user. Once a set of navigation sequences is encoded to a set of user embeddings, a variety of classifiers (prediction models) may be applied to the user embeddings to predict what a probable next-user-action may be and/or the likelihood that the next-user-action will be a desired target action.

Claims

1. A computer-implemented method for predicting a next action of a user, the method comprising:

obtaining a plurality of navigation sequences performed by the user;

applying a Recurrent Neural Network (RNN) to each navigation sequence of the plurality of navigation sequences to encode each navigation sequence to a user-embedding of a plurality of user-embedding; and

applying a classifier to each user-embedding to predict the next action of the user.

2. The computer-implemented method of claim 1, wherein each navigation sequence of the plurality of navigation sequences includes a set of user actions sequentially performed during a navigation session.

3. The computer-implemented method of claim 1, wherein each user-embedding reflects a temporally-defined navigation pattern of the user.

4. The computer-implemented method of claim 1, wherein the RNN is a Time-Aware RNN.

5. The computer-implemented method of claim 1, wherein the RNN is a Time-Aware Bi-Directional RNN.

6. The computer-implemented method of claim 1, wherein the RNN includes Long Short Term Memory architecture.

7. The computer-implemented method of claim 1, wherein each navigation sequence of the plurality of navigation sequences is obtained as a series of tuples, each tuple comprising (1) an identifier of an action taken by the user at a particular point in time, (2) an identifier associated with the user, and (3) an indicator of an amount of time spent on the action taken.

8. The computer-implemented method of claim 7, further comprising:

obtaining one of the plurality of navigation sequences performed by the user;

determining one of the plurality of user-embeddings is sufficient to predict a probable next-action of the user; and

applying the one of the plurality of user-embeddings to predict the probable next-action of the user.

9. The computer-implemented method of claim 7, further comprising:

obtaining one of the plurality of navigation sequences performed by the user;

determining that none of the plurality of user-embeddings is sufficient to predict a probable next-action of the user; and

applying one or more user embeddings associated with users other than the user to predict the probable next-action of the user.

10. A computer system comprising:

one or more processors; and

one or more computer storage media storing computer useable instructions to cause the one or more processors to;

obtain a set of navigation sequences associated with a set of users, each navigation sequence in the set of navigation sequences including a set of user actions sequentially performed during a navigation session, each navigation sequence being associated with a user included in the set of users,

apply a Recurrent Neural Network (RNN) to the set of navigation sequences to encode each navigation sequence in the set of navigation sequences into a user embedding; and

apply a classifier trained to a specific target action to the user embeddings to create a next-user-action predictive model for predicting the probability that a next-action performed by a target user is the target action.

11. The computer system of claim 10, wherein each navigation sequence is obtained as a series of tuples, each tuple comprising (1) an identifier of a user action taken at a particular point in time, (2) an identifier associated with a user that performed the user action taken, and (3) an indicator of an amount of time spent on the user action taken.

12. The computer system of claim 10, wherein each user-embedding reflects a temporally-defined navigation pattern of the associated user.

13. The computer system of claim 12, wherein the one or more processors are further caused to identify subsets of users, each subset including users included in the set of users that are associated with a similar temporally-defined navigation pattern.

14. The computer system of claim 13, wherein the one or more processors are caused to identify the subsets of users utilizing K-means clustering.

15. The computer system of claim 13, wherein the one or more processors are further caused to:

obtain a navigation sequence from a first user;

determine that one or more of the user embeddings associated with the first user are sufficient to predict a probability that a next-action performed by the first user is the target action; and

apply the next-user-action predictive model to the one or more user embeddings associated with the first user to predict a probability that the next-action performed by the first user is the target action.

16. The computer system of claim 13, wherein the one or more processors further are caused to:

obtain a navigation sequence from a first user;

determine that one or more of the user embeddings associated with the first user are insufficient to predict a probability that a next-action performed by the first user is the target action; and

apply the next-user-action predictive model to the one or more user embeddings associated with the first user and to user embeddings associated with other users included in a subset of users with which the first user has been identified to predict a probability that the next-action performed by the first user is the target action.

20. The computer-implemented method of claim 16, wherein the classifier is a target-action classifier that predicts a probability that a next-action performed by a target user is a specific target-action.

17. A computer-implemented method for generating next user-action predictive models using navigation sequences, the method comprising:

means for obtaining a set of navigation sequences associated with a set of users, each navigation sequence in the set of navigation sequences including a set of user actions performed during a navigation session;

means for applying a Recurrent Neural Network (RNN) to the set of navigation sequences to encode each navigation sequence in the set of navigation sequences into a user embedding; and

means for applying a classifier to the user embeddings to create a next-user-action predictive model for predicting next-actions of users.

18. The computer-implemented method of claim 17, wherein each navigation sequence being associated with a user included in the set of users, and each navigation sequence containing time-based information for each action in the set of user actions.

19. The computer-implemented method of claim 18, wherein each navigation sequence reflects a temporally-defined navigation pattern for the associated user, the RNN being a Time-Aware RNN and including Long Short Term Memory architecture.