Algorithmic apparel recommendation

Patent metadata

Field	Value
Publication number	US-2020257976-A1
Application number	US-202016778522-A
Country	US
Kind code	A1
Filing date	Jan 31, 2020
Priority date	Feb 7, 2019
Publication date	Aug 13, 2020
Grant date	—

Abstract

Generally, the present disclosure relates to methods and systems for algorithmically generating apparel recommendations. In some example aspects, human-identified complementarity of a subset of products can be used to train a neural network, which is in turn used to generate a compatibility score for items. Based on such values, compatible items can be identified and recommended to a user.

First claim

1 . A computer-implemented method comprising:
obtaining a pair of images of apparel items; providing the pair of images as input to a trained neural network; processing the pair of images with the trained neural network; obtaining a compatibility score as output from the trained neural network; and recommending an apparel item at a retail website based on the compatibility score, wherein processing the pair of images includes directly encoding correlation between embeddings.

2 . The method of claim 1 , wherein processing the pair of images with the trained neural network includes calculating a Hadamard product of the embeddings to directly encode correlation between the embeddings.

3 . The method of claim 1 , wherein processing the pair of images with the trained neural network includes incorporating color information into the trained neural network.

4 . The method of claim 3 , wherein incorporating color information into the neural network includes augmenting the embeddings with color histogram features obtained from the pair of images of apparel items.

5 . The method of claim 4 , wherein augmenting the embeddings includes concatenating the color histogram features with a Hadamard product of the embeddings.

6 . The method of claim 1 , wherein processing the pair of images with the trained neural network includes incorporating apparel category information into the trained neural network.

7 . The method of claim 6 , wherein the apparel category information is incorporated as embeddings of the pair of categories of the apparel items of the pair of images.

8 . A non-transitory computer-readable medium having stored thereon a neural network configured receive input representative of a pair of apparel images and provide an output representative of a compatibility score, the neural network comprising:
a trained first subnetwork configured to provide a pair of features as output, the trained first subnetwork comprising:
a left branch configured to generate embeddings for a first image of the pair of apparel images; and
a right branch configured to generate embeddings for a second image of the pair of apparel images;
a combiner configured to produce a vector from the pair of features; a second subnetwork configured to forward propagate the vector; and a readout function configured to produce the output representative of the compatibility score based on an output of the second subnetwork.

9 . The non-transitory computer-readable medium of claim 8 , wherein the trained first subnetwork is a siamese network.

10 . The non-transitory computer-readable medium of claim 8 , wherein the combiner is configured to calculate a Hadamard product of the embeddings for the first image and the embeddings for the second image.

11 . The non-transitory computer-readable medium of claim 10 , wherein the combiner is further configured to concatenate the Hadamard product with color histogram features extracted from the pair of apparel images.

12 . A computer-implemented method comprising:
receiving a seed item; for each respective item of a plurality of items in an item collection, determining a compatibility score between the seed item and the respective item, wherein the determining includes:
providing the seed item and the respective item as a pair of images as input to a trained neural network;
processing the pair of images with the trained neural network, wherein processing the pair of images includes directly encoding correlation between embeddings; and
obtaining the compatibility score as output from the trained neural network; and
providing a subset of the plurality of items recommending an apparel item at a retail website based on the compatibility score.

13 . The method of claim 12 , wherein processing the pair of images with the trained neural network includes calculating a Hadamard product of the embeddings to directly encode correlation between the embeddings.

14 . The method of claim 12 , wherein processing the pair of images with the trained neural network includes incorporating color information into the trained neural network.

15 . The method of claim 4 , wherein incorporating color information into the neural network includes augmenting the embeddings with color histogram features obtained from the pair of images of apparel items.

16 . The method of claim 5 , wherein augmenting the embeddings includes concatenating the color histogram features with a Hadamard product of the embeddings.

17 . The method of claim 12 , wherein processing the pair of images with the trained neural network includes incorporating apparel category information into the trained neural network.

18 . The method of claim 17 , wherein the apparel category information is incorporated as embeddings of the pair of categories of the apparel items of the pair.

19 . The method of claim 12 , wherein receiving the seed item includes receiving the seed item over the retail website.

20 . The method of claim 12 , wherein the trained neural network comprises:
a trained first subnetwork configured to provide a pair of features as output, the trained first subnetwork comprising:
a left branch configured to generate embeddings for a first image of the pair of images; and
a right branch configured to generate embeddings for a second image of the pair of images;
a combiner configured to produce a vector from the pair of features; a second subnetwork configured to forward propagate the vector; and a readout function configured to produce the output representative of the compatibility score based on an output of the second subnetwork.

Assignees

• Target Brands Inc

Inventors

• Polanía Cabrera Luisa F
• Gupte Satyajit Vishram

Classifications

• G06N3/044

  Recurrent networks, e.g. Hopfield networks · CPC title

• G06N3/048

  Activation functions · CPC title

• G06N3/045

  Combinations of networks · CPC title

• G06N3/09

  Supervised learning · CPC title

• G06N3/0464

  Convolutional networks [CNN, ConvNet] · CPC title