# US 2025/0061506 A1 · Systems and methods for generating recommendations using contextual bandit models with non-linear oracles

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

- Number: US 2025/0061506 A1
- Document ID: US20250061506A1
- Status: Published application
- Kind: A1
- Application number: 18/450,931
- Filed: 2023-08-16
- Published: 2025-02-20
- Earliest claimed: 2023-08-16
- Assignees: Intuit Inc.
- Applicants: Intuit Inc.
- Inventors: Shankar SANKARARAMAN; Isaac STORCH
- Web page: https://patentgrep.com/patent/US20250061506A1

## CPC

- G06Q 30/0631: {Recommending goods or services}
- 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

## Abstract

Systems and methods are provided for generating recommendations using contextual bandit models with non-linear oracles.

## Claims

### Claim 1

A computing system for generating recommendations comprising:
a processor; and
a non-transitory computer-readable storage device storing computer-executable instructions, the instructions operable to cause the processor to perform operations comprising:
detecting a presence of a user on a content platform;
generating a recommendation for the user via a contextual bandit model and a non-linear machine learning model, the recommendation comprising one or more recommended items, the non-linear machine learning model operating as an oracle for the contextual bandit model and being trained to predict future rewards for the one or more recommended items; and
collecting a reward from an interaction between the user and the one or more recommended items to re-train the non-linear machine learning model.

### Claim 2 (depends on 1)

The computing system of claim 1, wherein the operations comprise:
concatenating user features associated with the user, item features associated with the one or more recommended items, and context features to generate a vector;
assigning the collected reward to the vector to generate a vector-reward pair; and
re-training, with the vector-reward pair, the non-linear machine learning model to predict future rewards for the one or more recommended items, wherein the training utilizes a plurality of weights of the user features, item features, and context features in a hyperparameter search space.

### Claim 3 (depends on 1)

The computing system of claim 1, wherein training the non-linear machine learning model comprises training a tree-based machine learning model or a deep learning model.

### Claim 4 (depends on 1)

The computing system of claim 1, wherein collecting the reward comprises at least one of collecting a binary reward, a multi-class reward, or a continuous reward.

### Claim 5 (depends on 1)

The computing system of claim 1, wherein training the non-linear machine learning model comprises calculating, for each hyperparameter selection, an average reward uniqueness indicator (average RUI) value based on cross-validation data.

### Claim 6 (depends on 4)

The computing system of claim 4, wherein training the non-linear machine learning model comprises constraining each average RUI value with a predefined threshold.

### Claim 7 (depends on 1)

The computing system of claim 1, wherein generating the recommendation for the user via the contextual bandit model is performed based on exploration or exploitation.

### Claim 8

A computer-implemented method, performed by at least one processor, for generating recommendations comprising:
detecting a presence of a user on a content platform;
generating a recommendation for the user via a contextual bandit model and a non-linear machine learning model, the recommendation comprising one or more recommended items, the non-linear machine learning model operating as an oracle for the contextual bandit model and being trained to predict future rewards for the one or more recommended items; and
collecting a reward from an interaction between the user and the one or more recommended items to re-train the non-linear machine learning model.

### Claim 9 (depends on 8)

The computer-implemented method of claim 8, wherein the operations comprise:
concatenating user features associated with the user, item features associated with the one or more recommended items, and context features to generate a vector;
assigning the collected reward to the vector to generate a vector-reward pair; and
re-training, with the vector-reward pair, the non-linear machine learning model to predict future rewards for the one or more recommended items, wherein the training utilizes a plurality of weights of the user features, item features, and context features in a hyperparameter search space.

### Claim 10 (depends on 8)

The computer-implemented method of claim 8, wherein training the non-linear machine learning model comprises training a tree-based machine learning model or a deep learning model.

### Claim 11 (depends on 8)

The computer-implemented method of claim 8, wherein collecting the reward comprises at least one of collecting a binary reward, a multi-class reward, or a continuous reward.

### Claim 12 (depends on 8)

The computer-implemented method of claim 8, wherein training the non-linear machine learning model comprises calculating, for each hyperparameter selection an average reward uniqueness indicator (average RUI) value based on cross-validation data.

### Claim 13 (depends on 12)

The computer-implemented method of claim 12, wherein training the non-linear machine learning model comprises constraining each average RUI value with a predefined threshold.

### Claim 14 (depends on 8)

The computer-implemented method of claim 8, wherein generating the recommendation for the user via the contextual bandit model is performed based on exploration or exploitation.

### Claim 15

A computer-implemented method, performed by at least one processor, for training a model to generate recommendations comprising:
collecting a reward from an interaction between a user and one or more recommended items, the one or more recommended items being generated by a contextual bandit model;
concatenating user features associated with the user, item features associated with the one or more recommended items, and context features to generate a vector;
assigning the collected reward to the vector to generate a vector-reward pair; and
training, with the vector-reward pair, a non-linear machine learning model to predict future rewards for the one or more recommended items, wherein the training utilizes a plurality of weights of the user features, item features, and context features in a hyperparameter search space and the non-linear machine learning model operates as an oracle for the contextual bandit model.

### Claim 16 (depends on 15)

The computer-implemented method of claim 15, wherein training the non-linear machine learning model comprises training a tree-based machine learning model or a deep learning model.

### Claim 17 (depends on 15)

The computer-implemented method of claim 15, wherein collecting the reward comprises at least one of collecting a binary reward, a multi-class reward, or a continuous reward.

### Claim 18 (depends on 15)

The computer-implemented method of claim 15, wherein training the non-linear machine learning model comprises calculating, for each hyperparameter selection, an average reward uniqueness indicator (average RUI) value based on cross-validation data.

### Claim 19 (depends on 18)

The computer-implemented method of claim 18, wherein training the non-linear machine learning model comprises constraining each average RUI value with a predefined threshold.

### Claim 20 (depends on 15)

The computing system of claim 15, wherein training the non-linear machine learning model comprises training a tree-based machine learning model or a deep learning model.
