Recommendation system using improved neural network

Patent metadata

Field	Value
Publication number	US-10635973-B1
Application number	US-201615195515-A
Country	US
Kind code	B1
Filing date	Jun 28, 2016
Priority date	Jun 28, 2016
Publication date	Apr 28, 2020
Grant date	Apr 28, 2020

Abstract

Techniques described herein are directed to improved artificial neural network machine learning techniques that may be employed with a recommendation system to provide predictions with improved accuracy. In some embodiments, item consumption events may be identified for a plurality of users. From these item consumption events, a set of inputs and a set of outputs may be generated according to a data split. In some embodiments, the set of outputs (and potentially the set of inputs) may include item consumption events that are weighted according to a time-decay function. Once a set of inputs and a set of outputs are identified, they may be used to train a prediction model using an artificial neural network. The prediction model may then be used to identify predictions for a specific user based on user-specific item consumption event data.

First claim

What is claimed is:

1. A computer-implemented method, comprising:
identifying item consumption data for a plurality of users, the item consumption data including a record of past item consumption events for individual users in the plurality of users;
creating a prediction model with the item consumption data at least in part by:
determining a data split to be used in determining a set of inputs and a set of outputs;
determining a set of inputs from the item consumption data based on the data split;
determining a set of outputs from the item consumption data based on the data split, a portion of the set of outputs corresponding to a subset of the set of inputs; and
training the prediction model, wherein training the prediction model includes at least:
(1) involves determining a cost function based at least in part on a time-decay function applied to the portion of the set of outputs;
(2) utilizing the determined set of inputs to predict the determined set of outputs; and
(3) minimizing an error associated with the cost function based at least in part on the utilizing;

receiving user-specific item consumption data; and
generating, based at least in part on the prediction model and the user-specific item consumption data, a user-specific item recommendation.

2. The computer-implemented method of claim 1 , wherein the data split is a date threshold.

3. The computer-implemented method of claim 1 , wherein training the prediction model based on the determined set of inputs and the determined set of outputs comprises using at least one machine learning technique.

4. The computer-implemented method of claim 3 , wherein the at least one machine learning technique comprises an artificial neural network.

5. The computer-implemented method of claim 4 , wherein the artificial neural network is an autoencoder neural network.

6. A system comprising:
a processor; and
a memory including instructions that, when executed with the processor, cause the system to, at least:
determine a set of past item consumption events;
identify a first subset of the set of past item consumption events to be used as input values;
identify a second subset of the set of past item consumption events to be used as output values, the second subset including a portion that overlaps with the first subset;
train a prediction model, wherein training the prediction model includes at least:
(1) involves determining a cost function based at least in part on a time-decay function applied to the portion of the second subset;
(2) utilizing the first subset corresponding to the input values to predict the second subset corresponding to the output values; and
(2) minimizing an error associated with the cost function based at least in part on the utilizing;

receive an indication of at least one item consumption event associated with a user; and
process the at least one item consumption event with the prediction model to generate predicted item consumption events specific to the user.

7. The system of claim 6 , wherein the first subset of the set of past item consumption events and the second subset of the set of past item consumption events are identified based on a floating date threshold.

8. The system of claim 6 , wherein the prediction model is retrained on a periodic basis.

9. The system of claim 8 , wherein retraining the prediction model comprises updating both the identified first subset of the set of past item consumption events and the second subset of the set of past item consumption events.

10. The system of claim 6 , wherein the item consumption events comprise transactions for items in an electronic catalog.

11. The system of claim 10 , wherein the transactions for items in an electronic catalog include purchases of one or more items by a plurality of users.

12. The system of claim 10 , wherein the transactions for items in an electronic catalog include interactions with one or more websites associated with items in the electronic catalog.

13. The system of claim 6 , wherein the at least one item consumption event associated with the user is a prediction of an item consumption event likely to be performed by the user.

14. A non-transitory computer readable medium storing specific computer-executable instructions that, when executed by a processor, cause a computer system to, at least:
identify a first set of consumption events within a data store that meet a first condition;
identify a second set of consumption events within a data store, the second set of consumption events comprising:
at least a portion of the first set of consumption events; and
a plurality of additional consumption events;

train a prediction model, wherein training the prediction model includes at least:
(1) determining a cost function based at least in part on a time-decay function applied to the portion of the first set of consumption events within the second set of consumption events;
(2) utilizing the first set of consumption events to predict the second set of consumption events;
(3) minimizing an error associated with the cost function based at least in part on the utilizing; and

upon receiving a set of consumption events associated with a user, generate predictive consumption events to be associated with the user based at least in part on the trained prediction model.

15. The computer readable medium of claim 14 , wherein generating predictive consumption events to be associated with the user is also based at least in part on at least one of item features or user features.

16. The computer readable medium of claim 14 , wherein the prediction model includes at least one activation function.

17. The computer readable medium of claim 14 , wherein training the prediction model uses at least one machine learning technique.

18. The computer readable medium of claim 17 , wherein the cost function includes an element weighing an individual consumption event inversely proportional to a popularity of that individual consumption event.

19. The computer readable medium of claim 14 , wherein the cost function assigns a weight to each item consumption event based on whether an item associated with the item consumption event was consumed.

20. The computer readable medium of claim 14 , wherein the first set of consumption events is subjected to a second time-decay function.

Assignees

• Amazon Tech Inc

Inventors

• Dirac Leo Parker
• Joseph Rejith George
• Mohan Vijai
• Rybakov Oleg

Classifications

• G06N20/00

  Machine learning · CPC title

• G06N5/022

  Knowledge engineering; Knowledge acquisition · CPC title

• H04L51/24

  Electricity · mapped topic

• G06N3/08 Primary

  Learning methods · CPC title

• H04L51/32

  Electricity · mapped topic