Unsupervised embeddings disentanglement using a gan for merchant recommendations

Patent metadata

Field	Value
Publication number	US-12175504-B2
Application number	US-202218085034-A
Country	US
Kind code	B2
Filing date	Dec 20, 2022
Priority date	Sep 3, 2019
Publication date	Dec 24, 2024
Grant date	Dec 24, 2024

Abstract

Embodiments for training a recommendation system to provide merchant recommendations comprise receiving, by a processor, raw merchant embeddings and raw user embeddings generated from payment transaction records, wherein the raw merchant embeddings include a plurality of embedded features. A generative adversarial network (GAN) is trained to generate modified merchant embeddings from the raw merchant embeddings, where the modified embeddings remove a location feature. Subsequent to training and responsive to receiving a request for merchant recommendations in the target location for the target user, the GAN and a trained preference model are used to generate a list of merchant rankings based on a new set of modified merchant embeddings, past preferences of a target user, and the target location to recommend merchants in the target location.

First claim

We claim:

1. A computer-implemented method for training a recommendation system to provide merchant recommendations to a target user, comprising:
receiving, by one or more processors, historical payment transaction records of transactions made between users and merchants;
generating, by the one or more processors, from the historical payment transaction records a training set of raw merchant embeddings that represent the merchants and raw user embeddings that represent the users within an embedding space using a word embedding process, wherein the raw merchant embeddings include a plurality of embedded features from the historical payment transaction records including a location feature, a price feature, and a number of transactions feature, wherein the raw merchant embeddings include source merchant embeddings from a home location of the target user and target merchant embeddings from a target location;
training, by the one or more processors, a generative adversarial network (GAN) to generate modified merchant embeddings from the raw merchant embeddings, where the modified merchant embeddings have an effect of the location feature removed, the GAN comprising a generator and a discriminator, the GAN operational for:
receiving, by the generator, the source merchant embeddings as a set X of n embeddings X=x1, . . . , xn from the home location;
receiving, by the discriminator, the target merchant embeddings as a set Y of m embeddings Y=y1, . . . , ym from the target location; and
generating, by the generator, the modified merchant embeddings by applying a mapping function F to the set X of the source merchant embeddings, F (X)=F (x1), . . . , F (xn), wherein the generator uses F (X) to map the modified merchant embeddings to an embeddings space such that the discriminator cannot tell a difference between F (x 1 ), . . . , F (x n ) and the set Y of the target merchant embeddings; and
discriminating, by the discriminator, between elements randomly sampled from F (X)=F (x1), . . . , F (xn) and Y to identify an origin of the modified merchant embeddings;

continue, by the one or more processors, training the GAN until the discriminator cannot tell a difference between the target merchant embeddings and the modified merchant embeddings, indicating the location feature in the source merchant embeddings is mapped to the target merchant embeddings; and
subsequent to the training and responsive to receiving a request for merchant recommendations in the target location for the target user, generating by the GAN a list of merchant rankings based on a new set of modified merchant embeddings generated from a new set of payment transaction records made between the users and merchants, past preferences of the target user, and the target location to recommend merchants in the target location.

2. The method of claim 1 , wherein the plurality of embedded features include a cuisine type feature.

3. The method of claim 1 , wherein generating a list of merchant rankings based on the new set of modified merchant embeddings further comprises:
receiving, by the one or more processors, the new set of payment transaction records of transactions made between the users and the merchants;
generating, by the one or more processors, from the new set of payment transaction records new raw merchant embeddings and new raw user embeddings;
performing, by the one or more processors, a domain-adversarial approach to map the plurality of embedded features from the source merchant embeddings to the target location together to a common embedding space using the GAN that performs a disentanglement process on the new raw merchant embeddings to remove the effect of the location feature by generating a new set of modified merchant embeddings that are free of the location feature;
using, by the one or more processors, a preference model to determine new past preferences of the target user for the merchants based on the new modified merchant embeddings and the new raw user embeddings;
automatically generating, by the one or more processors, the list of merchant rankings based on the new set of modified merchant embeddings, the new past preferences of the target user, and the target location to recommend merchants in the target location; and
providing, by the one or more processors, the list of merchant rankings to the target user.

4. The method of claim 3 , wherein the disentanglement process performed by the GAN further comprises:
receiving, by the generator, new source merchant embeddings represented as a new set X of n embeddings X=x1, . . . , xn from the source location comprising the home location of the target user; and
receiving, by the discriminator, new target merchant embeddings represented as a new set Y of m embeddings Y=y1, . . . , ym from a new target location comprising a current location from a new target user.

5. The method of claim 4 further comprising:
implementing the generator as a deep neural network comprising an input layer, one or more hidden layers and an output layer, wherein the input layer and the output layer contain a same number of nodes that match a dimensionality of the new source merchant embeddings and the new modified merchant embeddings; and
implementing the discriminator as a type of classifier deep neural network comprising an input layer, two or more hidden layers and an output layer, wherein each successive hidden layer contains less nodes than a previous hidden layer.

6. The method of claim 4 further comprising: performing the disentanglement on multiple locations by one of: i) a pairwise procedure that selects a first location and then maps the new raw merchant embeddings from remaining locations to an embedding space of the first location; and ii) mapping the new raw merchant embeddings from the multiple locations to a new embedding space.

7. The method of claim 4 further comprising: calculating, by the generator, a loss function L G and calculating, by the discriminator, a loss function L D , where:

L
G

=

1
n

⁢

∑
n

i
=
1

(

α
·

L
cos

(

F
⁡
(

X
i

)

,

X
i

)

+

β
·

L
sig

Assignees

• Visa Int Service Ass

Inventors

• Zheng Yan
• Wang Yuwei
• Zhang Wei
• Yeh Michael
• Wang Liang

Classifications

• G06N3/094

  Adversarial learning · CPC title

• G06N3/0895

  Weakly supervised learning, e.g. semi-supervised or self-supervised learning · CPC title

• G06N3/0475

  Generative networks · CPC title

• G06N3/09

  Supervised learning · CPC title

• G06Q30/0201

  Market modelling; Market analysis; Collecting market data · CPC title