Digital experience enhancement using an ensemble deep learning model

Patent metadata

Field	Value
Publication number	US-11816562-B2
Application number	US-201916375627-A
Country	US
Kind code	B2
Filing date	Apr 4, 2019
Priority date	Apr 4, 2019
Publication date	Nov 14, 2023
Grant date	Nov 14, 2023

Abstract

A digital experience enhancement system includes an ensemble deep learning model that includes an estimator ensemble and a neural network. The ensemble deep learning model is trained to generate a digital experience enhancement recommendation from an enhancement request. The ensemble deep learning model receives the enhancement request, which is input to the estimator ensemble. The estimator ensemble uses various different machine learning systems to generate estimator output values. The neural network uses the estimator output values from the estimator ensemble to generate a digital experience enhancement recommendation. The digital experience generation system then uses this digital experience enhancement recommendation to enhance the digital experience.

First claim

What is claimed is:

1. A method comprising:
generating, by a processing device, a trained estimator ensemble to generate multiple estimation values by training, using a first training data set including ground truths that are specific to a digital experience use scenario based on past user interactions with a digital experience displayed in a user interface, an estimator combination of a singular value decomposition estimator, a neighborhood or clustering estimator, a factorization estimator, a time-aware estimator that includes a time-aware neural factorization estimator leveraging a time that a value is provided by the user, a variational autoencoder estimator, and a gradient boosting estimator included in an estimator ensemble to generate an estimation value;
training, by the processing device using a second training data set and the multiple estimation values generated by the trained estimator ensemble, a neural network to generate recommendations to enhance the digital experience for the user, the neural network including nodes with weights that are tuned to minimize root mean square errors between the ground truths and the recommendations to enhance the digital experience for the user;
updating, by the processing device, the weights using stochastic gradient descent with Nesterov momentum and applying an Adam optimizer;
receiving, by the processing device, a request for a recommendation to enhance the digital experience for the user, the request including an indication of past user interactions of the user with the digital experience;
generating, by the processing device using the trained estimator ensemble and based on the indication of past user interactions, the multiple estimation values;
generating, by the processing device using the neural network and based on the multiple estimation values, the recommendation to enhance the digital experience for the user;
enhancing, by the processing device, the digital experience based on the recommendation to generate an enhanced digital experience; and
displaying, by the processing device, the enhanced digital experience.

2. The method of claim 1 , the recommendation being one of multiple potential values, the neural network comprising a 3-layer neural network followed by a mapping and normalization layer, the mapping and normalization layer outputting the recommendation as a set of probability distributions on the multiple values.

3. The method of claim 2 , the training the neural network including minimizing cross-entropy loss between the recommendations and one-hot representations of ground truths using a loss function and a regression tree.

4. The method of claim 1 , the recommendation being one of multiple potential values, the neural network comprising a 3-layer neural network outputting the recommendation as a single value output.

5. The method of claim 1 , the past user interactions including values provided by the user for different items included in the digital experience.

6. The method of claim 5 , the past user interactions further including a time feature that indicates, for a particular item, a time that the particular item was first available to the user.

7. The method of claim 5 , the past user interactions further including a time feature that indicates, for a particular item, a time that the user provided the value for the particular item.

8. The method of claim 5 , the past user interactions further including a time feature that indicates a time that the user first provided a value for any of the different items.

9. The method of claim 5 , the past user interactions further including a time feature that indicates, for a particular item, a timespan between a time that the particular item was first available to the user and a time that the user provided the value for the particular item.

10. The method of claim 1 , wherein the neighborhood or clustering estimator comprises an item k-nearest neighbors estimator that generates an estimator output value if a threshold confidence level is achieved.

11. The method of claim 10 , wherein the neighborhood or clustering estimator further comprises a neighboring average estimator that averages the estimator output value from the item k-nearest neighbors estimator with an output generated by a k-means estimator.

12. A method comprising:
obtaining, by a processing device, a first training data set that includes, for each of multiple users, values associated with the user for particular items including ground truths that are specific to a digital experience use scenario based on past user interactions with a digital experience displayed in a user interface;
training, by the processing device in a first stage using the first training data set, an estimator ensemble including an estimator combination of a singular value decomposition estimator, a neighborhood or clustering estimator, a factorization estimator, a time-aware estimator that includes a time-aware neural factorization estimator leveraging a time that a value is provided by the user, a variational autoencoder estimator, and a gradient boosting estimator in an estimator ensemble to generate an estimation value;
obtaining, by the processing device, a second training data set that includes, for each of the multiple users, values associated with the user for particular items;
generating, by the processing device using the estimator ensemble previously trained in the first stage and the second training data set, multiple estimation values;
training, by the processing device in a second stage using the multiple estimation values, a neural network to generate a recommendation to enhance the digital experience for the user, the neural network including nodes with weights that are tuned to minimize root mean square errors between the ground truths and the recommendation to enhance the digital experience for the user, the weights are updated using stochastic gradient descent with Nesterov momentum and applying an Adam optimizer; and
enhancing, by the processing device using the recommendation, the digital experience for the user.

13. The method of claim 12 , the recommendation being one of multiple potential values, the neural network comprising a 3-layer neural network followed by a mapping and normalization layer, the mapping and normalization layer outputting the recommendation as a set of probability distributions on the multiple values, and the training the neural network comprising training the neural network to minimize cross-entropy loss between the recommendation and a one-hot representation of a ground truth using a loss function and a regression tree.

14. The method of claim 12 , the recommendation being one of multiple potential values, the neural network comprising a 3-layer neural network outputting the recommendation as a single value output.

15. The method of claim 12 , the particular items including movies.

16. The method of claim 12 , wherein the first training data set and the second training data set are two different training data sets.

17. A system comprising:
means for generating a trained estimator ensemble by training, using a first training data set including ground truths that are specific to a digital experience use scenario based on past user interactions with a digital experience displayed in a user interface, an estimator ensemble including an estimator combination of a singular value decomposition estimator, a neighborhood or clustering estimator, a factorization estimator, a time-aware estimator that includes a time-aware neural factorization estimator leveraging a time that a value is provi

Assignees

• Adobe Inc

Inventors

• Burkhart Michael Craig
• Modarresi Kourosh

Classifications

• G06N3/09

  Supervised learning · CPC title

• G06N3/0455

  Auto-encoder networks; Encoder-decoder networks · CPC title

• G06N3/0499

  Feedforward networks · CPC title

• G06N3/08 Primary

  Learning methods · CPC title

• G06N20/20 Primary

  Ensemble learning · CPC title