# US 11,443,346 B2 · Group item recommendations for ephemeral groups based on mutual information maximization

> Agents: this data is available as JSON from the patentgrep API and MCP server — see https://patentgrep.com/llms.txt

- Number: US 11,443,346 B2
- Document ID: US11443346B2
- Status: Granted
- Kind: B2
- Application number: 17/070,772
- Filed: 2020-10-14
- Published: 2021-04-15
- Granted: 2022-09-13
- Earliest claimed: 2019-10-14
- Assignees: Visa International Service Association
- Applicants: Visa International Service Association
- Inventors: Aravind Sankar; Yanhong Wu; Yuhang Wu; Wei Zhang; Hao Yang
- Web page: https://patentgrep.com/patent/US11443346B2

## CPC

- G06Q 30/0269: {based on user profile or attribute}
- G06F 16/335: Filtering based on additional data, e.g. user or group profiles (filtering in web context G06F16/9535, G06F16/9536)
- G06N 3/045: Combinations of networks
- G06N 3/0455: Auto-encoder networks; Encoder-decoder networks
- G06N 3/0499: Feedforward networks
- G06N 3/08: Learning methods
- G06N 3/09: Supervised learning
- G06N 7/01: Probabilistic graphical models, e.g. probabilistic networks
- G06Q 30/0251: Targeted advertisements
- H04L 67/306: User profiles
- H04L 67/535: {Tracking the activity of the user (network monitoring arrangements H04L43/00; recording of computer activity G06F11/34)}
- G06Q: INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06F: ELECTRIC DIGITAL DATA PROCESSING (computer systems based on specific computational models G06N)
- G06N: COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- H04L: TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION (arrangements common to telegraphic and telephonic communication H04M)

## Abstract

A computer-implemented method is disclosed for training neural networks of a group recommender to provide item recommendations for ephemeral groups having group interaction sparsity. A preference encoder and aggregator generate user and group preference embeddings from user-item interactions, wherein the preference embeddings form a latent user-group latent embedding space. The neural preference encoder and the aggregator are trained by regularizing the latent user-group embedding space to overcome the group interaction sparsity by: i) maximizing user-group mutual information (MI) between the group embeddings and the user embeddings so that the group embeddings encode shared group member preferences, while regularizing the user embeddings to capture user social associations, and ii) contextually identifying informative group members and regularizing the corresponding group embeddings using a contextually weighted user loss value to contextually weight users' personal preferences in proportion to their user-group MI to reflect personal preferences of the identified informative group members.

## Claims

### Claim 1

A computer-implemented method of training embedding neural networks of a base group recommender to provide item recommendations for ephemeral groups of group members, the ephemeral groups having group interaction sparsity, the method comprising:
providing a neural preference encoder to generate user preference embeddings from individual user-item interactions extracted from payment transaction records stored in a transaction database;
providing a neural preference aggregator to aggregate the user preference embeddings into group preference embeddings, wherein the user preference embeddings and the group preference embeddings form a latent user-group latent embedding space; and
training the neural preference encoder and the neural preference aggregator by regularizing the latent user-group embedding space to overcome the group interaction sparsity by:
i) training a discriminator to maximize user-group mutual information (MI) between the group preference embeddings and the user preference embeddings of the group members so that the group preference embeddings encode shared group member preferences, while regularizing the user preference embeddings to capture user social associations, and
ii) contextually identifying informative group members and regularizing the corresponding group preference embeddings using a contextually weighted user loss (LUG) value based at least in part on the output of the discriminator to contextually weight users' personal preferences in each group in proportion to their user-group MI to reflect personal preferences of the identified informative group members.

### Claim 2 (depends on 1)

The method of claim 1, wherein training the neural preference encoder and the neural preference aggregator further comprises: configuring the training process to train arbitrary types neural preference encoders and aggregators for ephemeral group recommendation to provide a recommender architecture-agnostic framework.

### Claim 3 (depends on 1)

The method of claim 1, wherein training the neural preference encoder and the neural preference aggregator further comprises: accessing the payment transaction records from the transaction database and extracting user, group, and item information including a training user set U and a training group set G, and for each user and group, determine individual user-item interactions xu and individual group-item interactions xg, respectively.

### Claim 4 (depends on 3)

The method of claim 3, further comprising: training the neural preference encoder, the neural preference aggregator, and learnable weights W used therein, on the individual user-item interactions xu to generate user preference embeddings eu and group preference embeddings eg, respectively, using a group loss (LG) value, a user-group mutual information (MI) maximization (LMI) value and a contextually weighted user loss (LUG) value.

### Claim 5 (depends on 4)

The method of claim 4, wherein training the neural preference encoder and the neural preference aggregator further comprises: saving for each of the groups the group preference embedding eg and the individual group-item interactions xg as a feature pair in a feature database.

### Claim 6 (depends on 5)

The method of claim 5, wherein the neural preference encoder and the neural preference aggregator comprise components a neural group recommender, the method further comprising: operating the neural group recommender in an operational mode comprising:
receiving user, group, and item information, including a user set U* and a group set G*, and for each user determining individual user-item interactions xu*;
using the neural preference encoder to predict user preference embeddings eu* based on the individual user-item interactions xu*; using the neural preference aggregator to aggregate the user preference embeddings eu* into group preference embeddings eg*;
retrieving from the feature database, by the neural group recommender, the feature pair having the group preference embedding eg that is most similar to the group preference embeddings eg* and setting the individual group-item interactions xg* equal to the individual group-item interactions xg from the feature pair; and
sending the individual group-item interactions xg* to a user of group gas the ranked list of items.

### Claim 7 (depends on 4)

The method of claim 4, further comprising: using the user-group MI to measure contextual informativeness of a group member u towards the group decision through the reduction in group decision uncertainty when user u is included in group g.

### Claim 8 (depends on 7)

The method of claim 7, further comprising: training the discriminator D to accurately separate positive samples drawn from a joint distribution from negative samples drawn from the product of marginals.

### Claim 9 (depends on 8)

The method of claim 8, further comprising: using as the positive samples (eu, eg) for D the preference embeddings of (u, g) pairs such that u∈ug, and deriving the negative samples by pairing eg with the preference embeddings of non-member users sampled from a negative sampling distribution, wherein the negative sample encourage the discriminator to learn latent aspects shared by group members in contrast to other users with similar personal item histories.

### Claim 10 (depends on 8)

The method of claim 8, further comprising: outputting, by the discriminator, a probability score D(eu, eg) assigned to user-group pairs, where eu representing user preference embeddings and eg representing group preference embeddings, where a higher score is assigned for users who are members of the group.

### Claim 11

A neural group recommender system, comprising:
a transaction database storing payment transaction records;
a processor having access to the transaction database; and
a software component executed by the processor, the software component comprising:
a neural preference encoder to generate user preference embeddings from individual user-item interactions extracted from payment transaction records stored in a transaction database;
a neural preference aggregator to aggregate the user preference embeddings into group preference embeddings, wherein the user preference embeddings and the group preference embeddings form a latent user-group latent embedding space;
wherein the software component is configured to train the neural preference encoder and the neural preference aggregator by regularizing the latent user-group embedding space to overcome the group interaction sparsity by:
i) training a discriminator to maximize user-group mutual information (MI) between the group preference embeddings and the user preference embeddings of the group members so that the group preference embeddings encode shared group member preferences, while regularizing the user preference embeddings to capture user social associations, and
ii) contextually identifying informative group members and regularizing the corresponding group preference embeddings using a contextually weighted user loss (LUG) value based at least in part on the output of the discriminator to contextually weight users' personal preferences in each group in proportion to their user-group MI to reflect personal preferences of the identified informative group members.

### Claim 12 (depends on 11)

The neural group recommender system of claim 11, wherein the training is for configured to train arbitrary types neural preference encoders and aggregators for ephemeral group recommendation to provide a recommender architecture-agnostic framework.

### Claim 13 (depends on 11)

The neural group recommender system of claim 11, wherein the payment transaction records are accessed from the transaction database to extract user, group, and item information including a training user set U and a training group set G, and for each user and group, determine individual user-item interactions xu and individual group-item interactions xg, respectively.

### Claim 14 (depends on 13)

The neural group recommender system of claim 13, wherein the neural preference encoder, the neural preference aggregator, and learnable weights W used therein, are trained on the individual user-item interactions xu to generate user preference embeddings eu and group preference embeddings eg, respectively, using a group loss (LG) value, a user-group mutual information (MI) maximization (LMI) value and a contextually weighted user loss (LUG) value.

### Claim 15 (depends on 14)

The neural group recommender system of claim 14, wherein the group preference embedding eg and the individual group-item interactions xg as a feature pair in a feature database for each of the groups.

### Claim 16 (depends on 15)

The neural group recommender system of claim 15, wherein the neural preference encoder and the neural preference aggregator comprise components a neural group recommender, the neural group recommender system operating in an operational mode to:
receive user, group, and item information, including a user set U* and a group set G*, and for each user determining individual user-item interactions xu*;
use the neural preference encoder to predict user preference embeddings eu* based on the individual user-item interactions xu*; using the neural preference aggregator to aggregate the user preference embeddings eu* into group preference embeddings eg*;
retrieve from the feature database, by the neural group recommender, the feature pair having the group preference embedding eg that is most similar to the group preference embeddings eg* and setting the individual group-item interactions xg* equal to the individual group-item interactions xg from the feature pair; and
send the individual group-item interactions xg* to a user of group g as the ranked list of items.

### Claim 17 (depends on 15)

The neural group recommender system of claim 15, wherein the user-group MI is used to measure contextual informativeness of a group member u towards the group decision through the reduction in group decision uncertainty when user u is included in group g.

### Claim 18 (depends on 17)

The neural group recommender system of claim 17, wherein the discriminator D is trained to accurately separate positive samples drawn from a joint distribution from negative samples drawn from the product of marginals.

### Claim 19 (depends on 18)

The neural group recommender system of claim 18, wherein the preference embeddings of (u, g) pairs are used as the positive samples (eu, eg) for D such that u∈ug, and the negative samples are derived by pairing eg with the preference embeddings of non-member users sampled from a negative sampling distribution, wherein the negative sample encourage the discriminator to learn latent aspects shared by group members in contrast to other users with similar personal item histories.

### Claim 20

A non-transitory computer-readable medium containing program instructions, which when executed by a processor cause the processor to train embedding neural networks of a base group recommender to provide item recommendations for ephemeral groups of group members, the ephemeral groups having group interaction sparsity, the program instructions for:
providing a neural preference encoder to generate user preference embeddings from individual user-item interactions extracted from payment transaction records stored in a transaction database;
providing a neural preference aggregator to aggregate the user preference embeddings into group preference embeddings, wherein the user preference embeddings and the group preference embeddings form a latent user-group latent embedding space; and
training the neural preference encoder and the neural preference aggregator by regularizing the latent user-group embedding space to overcome the group interaction sparsity by:
i) training a discriminator to maximize user-group mutual information (MI) between the group preference embeddings and the user preference embeddings of the group members so that the group preference embeddings encode shared group member preferences, while regularizing the user preference embeddings to capture user social associations, and
ii) contextually identifying informative group members and regularizing the corresponding group preference embeddings using a contextually weighted user loss (LUG) value based at least in part on the output of the discriminator to contextually weight users' personal preferences in each group in proportion to their user-group MI to reflect personal preferences of the identified informative group members.