Complementary-similarity learning using a quadruplet network

Patent metadata

Field	Value
Publication number	US-11416908-B2
Application number	US-202016779133-A
Country	US
Kind code	B2
Filing date	Jan 31, 2020
Priority date	Aug 23, 2019
Publication date	Aug 16, 2022
Grant date	Aug 16, 2022

Abstract

A system including one or more processors and one or more non-transitory computer-readable media storing computing instructions configured to run on the one or more processors and perform generating a training dataset comprising training quadruplets; generating a respective text feature vector for each of the four respective items for the each of the training quadruplets using a vector encoder; transforming the respective text feature vector for each of the four respective items; training the shared trainable parameters of the feature representation transformation model; receiving, from a user device a selection of an anchor item from the item catalog; determining, for the anchor item, one or more similar items or one or more complementary items; and sending instructions to display the one or more of the one or more similar items or the one or more of the one or more complementary items on the user device. Other embodiments are disclosed.

First claim

What is claimed:

1. A system comprising:
one or more processors; and
one or more non-transitory computer-readable media storing computing instructions configured to run on the one or more processors and perform:
generating a training dataset comprising training quadruplets, wherein each of the training quadruplets comprises four respective items of an item catalog, the four respective items comprising (a) a respective anchor item, (b) a respective similar item, (c) a respective complementary item, and (d) a respective negative item;
generating a respective text feature vector for each of the four respective items for the each of the training quadruplets using a vector encoder with shared non-trainable parameters across the four respective items;
transforming the respective text feature vector for each of the four respective items for the each of the training quadruplets into a respective transformed representation using a feature representation transformation model having shared trainable parameters across the four respective items;
training the shared trainable parameters of the feature representation transformation model, for the each of the training quadruplets, using a quadruplet loss for the four respective items of the each of the training quadruplets, to differentiate between the respective transformed representations for the four respective items;
receiving, from a user device, a selection of an anchor item from the item catalog;
determining, for the anchor item, one or more of (a) one or more similar items from the item catalog or (b) one or more complementary items from the item catalog, using the feature representation transformation model, as trained; and
sending instructions to display on the user device the one or more of (a) the one or more similar items or (b) the one or more complementary items.

2. The system of claim 1 , wherein the transforming the respective text feature vector using a feature representation transformation model further comprises:
using a neural network comprising a first fully connected layer with a Rectified Linear Unit (ReLU) and a second fully connected layer.

3. The system of claim 2 , wherein:
a dimension of a vector embedding output of the first fully connected layer with the ReLU is 256; and
a dimension of the respective transformed representation for each of the four respective items for the each of the training quadruplets generated by the second fully connected layer is 128.

4. The system of claim 1 , wherein:
a dimension of the respective text feature vector generated by the vector encoder is 512; and
the vector encoder is a Google Universal Sentence Encoder (“GUSE”).

5. The system of claim 1 , wherein training the shared trainable parameters of the feature representation transformation model further comprises:
determining the quadruplet loss, for each of the training quadruplets, based on a quadruplet loss function comprising a similarity loss function, a complementary loss function, and a negative loss function.

6. The system of claim 5 , wherein determining the quadruplet loss further comprises:
generating the similarity loss function using a similarity margin.

7. The system of claim 6 , wherein determining the quadruplet loss further comprises:
generating the complementary loss function using the similarity margin and a complementary margin.

8. The system of claim 5 , wherein determining the quadruplet loss further comprises:
generating the negative loss function using a negative margin.

9. The system of claim 1 , wherein the one or more similar items to the anchor item are determined based on respective second transformed representations generated using the feature representation model, as trained, for each of the one or more similar items in the item catalog being within a similarity margin.

10. The system of claim 1 , wherein the one or more complementary items to the anchor item are determined based on respective second transformed representations generated using the feature representation model, as trained, for each of the one or more complementary items in the item catalog being outside a similarity margin and inside a complementary margin.

11. A method being implemented via execution of computing instructions configured to run at one or more processors and stored at one or more non-transitory computer-readable media, the method comprising:
generating a training dataset comprising training quadruplets, wherein each of the training quadruplets comprises four respective items of an item catalog, the four respective items comprising (a) a respective anchor item, (b) a respective similar item, (c) a respective complementary item, and (d) a respective negative item;
generating a respective text feature vector for each of the four respective items for the each of the training quadruplets using a vector encoder with shared non-trainable parameters across the four respective items;
transforming the respective text feature vector for each of the four respective items for the each of the training quadruplets into a respective transformed representation using a feature representation transformation model having shared trainable parameters across the four respective items;
training the shared trainable parameters of the feature representation transformation model, for the each of the training quadruplets, using a quadruplet loss for the four respective items of the each of the training quadruplets, to differentiate between the respective transformed representations for the four respective items;
receiving, from a user device, a selection of an anchor item from the item catalog;
determining, for the anchor item, one or more of (a) one or more similar items from the item catalog or (b) one or more complementary items from the item catalog, using the feature representation transformation model, as trained; and
sending instructions to display on the user device the one or more of (a) the one or more similar items or (b) the one or more complementary items.

12. The method of claim 11 , wherein the transforming the respective text feature vector using a feature representation transformation model further comprises:
using a neural network comprising a first fully connected layer with a Rectified Linear Unit (ReLU) and a second fully connected layer.

13. The method of claim 12 , wherein:
a dimension of a vector embedding output of the first fully connected layer with the ReLU is 256; and
a dimension of the respective transformed representation for each of the four respective items for the each of the training quadruplets generated by the second fully connected layer is 128.

14. The method of claim 11 , wherein:
a dimension of the respective text feature vector generated by the vector encoder is 512; and
the vector encoder is a Google Universal Sentence Encoder (“GUSE”).

15. The method of claim 11 , wherein training the shared trainable parameters of the feature representation transformation model further comprises:
determining the quadruplet loss, for each of the training quadruplets, based on a quadruplet loss function comprising a similarity loss function, a complementary loss function, and a negative loss function.

16. The method of claim 15 , wherein determining the quadruplet loss further comprises:
generating the similarity loss function using a similarity margin.

17. The method of claim 16 , wherein determining the quadruplet loss further comprises:
generating the complementary loss function using the similarity margin and a complementary margin.

18. The method

Assignees

• Walmart Apollo Llc

Inventors

• Mane Mansi Ranjit
• Guo Stephen Dean
• Achan Kannan

Classifications

• G06N3/048

  Activation functions · CPC title

• G06N3/045

  Combinations of networks · CPC title

• G06N3/0499

  Feedforward networks · CPC title

• G06N3/09

  Supervised learning · CPC title

• G06N3/08

  Learning methods · CPC title