Recommendation system using improved neural network

Patent metadata

Field	Value
Publication number	US-10650432-B1
Application number	US-201615362585-A
Country	US
Kind code	B1
Filing date	Nov 28, 2016
Priority date	Nov 28, 2016
Publication date	May 12, 2020
Grant date	May 12, 2020

Abstract

Some aspects of the present disclosure relate to generating and training a neural network by separating historical item interaction data into both inputs and outputs. This may be done, for example, based on date. For example, a neural network machine learning technique may be used to generate a prediction model using a set of inputs that includes both a number of items purchased by a number of users before a certain date as well as some or all attributes of those items, and a set of outputs that includes the items purchased after that date. The items purchased before that date and the associated attributes can be subjected to a time-decay function.

First claim

What is claimed is:

1. A non-transitory computer readable medium storing a computer-executable module that, when executed by a computing system comprising one or more processors, causes the computing system to perform a process comprising:
obtaining first training data representing a plurality of user-item interaction events during a predetermined timeframe, each user-item interaction event comprising an interaction between a human user of an electronic catalog and a representation of one of a plurality of items in the electronic catalog;
obtaining second training data representing a plurality of attributes associated with each of the plurality of user-item interaction events;
assigning, to an input group:
a first subset of the plurality of items associated with a less recent subset of the plurality of user-item interaction events occurring within a first window between a beginning of the predetermined timeframe and a data split time, and
a subset of the plurality of attributes associated with the less recent subset of the plurality of user-item interaction events, and

assigning, to an expected output group, a second subset of the plurality of items associated with user-item interaction events occurring within a second window between the data split time and an end of the predetermined timeframe;
training a machine learning model to predict the expected output group based on the input group, wherein training the machine learning model comprises (1) providing the input group to input layer nodes of a neural network, and (2) providing the expected output group to output layer nodes of the neural network;
obtaining data representing at least a first item and a second item available in the electronic catalog; and
generating, using the machine learning model and the data representing the first item and the plurality of other items available in the electronic catalog, a score reflecting a likelihood that a user interacting with the first item will also interact with the second item.

2. The non-transitory computer readable medium of claim 1 , wherein the user-item interaction events comprise one or more of an item purchase, item view, and item share event.

3. The non-transitory computer readable medium of claim 1 , wherein the first training data includes sets of histories of a plurality of different users of the electronic catalog, and wherein the input group and output group are populated with items and associated attributes only from sets of histories including data in both the first and second windows.

4. The non-transitory computer readable medium of claim 1 , wherein the process performed by the computer-executable module, when executed by the computing system, further comprises:
receiving an indication of user interaction with a representation of the first item in the electronic catalog; and
causing display of a recommendation for the second item.

5. The non-transitory computer readable medium of claim 4 , wherein the indication comprises a user viewing a network page including information representing the first item or queuing the first item for purchase.

6. The non-transitory computer readable medium of claim 1 , wherein the first window is substantially shorter in time than the second window.

7. The non-transitory computer readable medium of claim 1 , wherein the computer-executable module, when executed, causes some user-item interaction events in a user's event history to be allocated solely to the input group and causes other user-item interaction events in the user's event history to be allocated solely to the expected output group.

8. A computer-implemented method, comprising:
maintaining user-specific event histories for each of a plurality of users of an electronic catalog, each event history corresponding to a respective user and comprising event data representing user-item interaction events performed by the user, wherein the event data for each user-item interaction event includes an identifier of a catalog item corresponding to the event and a time of the user-item interaction;
dividing each user-specific event history such that user-item interaction events that occurred before a split time are allocated exclusively to an input group and user-item interaction events that occurred after the split time are allocated exclusively to an expected output group;
training a neural network to predict user-item interaction events falling in the expected output group, wherein training the neural network comprises (1) providing, to input layer nodes of the neural network, identifiers and attributes of catalog items represented in the input group, and (2) providing, to output layer nodes of the neural network, identifiers of catalog items represented in the expected output group; and
generating, with the trained neural network, a score representing a likelihood that a user that interacts with a particular first catalog item will also interact with a particular second catalog item;
said method performed by execution of program instructions by a computing system.

9. The method of claim 8 , wherein the input group corresponds to a time window that is substantially shorter in duration than a time window corresponding to the expected output group.

10. The method of claim 8 , wherein training the neural network further comprises providing, to the input layer nodes, user attributes of users corresponding to the user-specific event histories.

Assignees

• Amazon Tech Inc

Inventors

• Joseph Rejith George
• Rybakov Oleg

Classifications

• G06Q30/0631 Primary

  Recommending goods or services · CPC title

• G06N3/08

  Learning methods · CPC title

• G06N3/09

  Supervised learning · CPC title

• G06N3/0455

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

• G06N3/0499

  Feedforward networks · CPC title