# US 2025/0156751 A1 · Real-time personalized recommendations by an ultrafast, lightweight, highly performant system

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

- Number: US 2025/0156751 A1
- Document ID: US20250156751A1
- Status: Published application
- Kind: A1
- Application number: 18/509,004
- Filed: 2023-11-14
- Published: 2025-05-15
- Earliest claimed: 2023-11-14
- Assignees: Adobe Inc.
- Applicants: Adobe Inc.
- Inventors: Nikolaos VLASSIS; Oliver BRDICZKA; Nipun JINDAL; Marcos Berteli SLOMP; Kevin John WAMPLER; Anand KHANNA; Amine Ben KHALIFA
- Web page: https://patentgrep.com/patent/US20250156751A1

## CPC

- G06N 20/00: Machine learning
- G06N: COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS

## Abstract

Users interact with items, such as movies, music, and document templates, among others. Item recommendations based on these user interactions are determined and provided to a user. A binary matrix indicating what items users have interacted with is provided. A design matrix is determined from the binary matrix. In this format, the model can be processed in parallel by a computing device. Columns of the design matrix are processed by threads of one or more CPUs of a computing system, in which a least squares analysis is performed over each thread. The output of the processing is a trained model useable for outputting an item recommendation responsive to an input user interaction during inference.

## Claims

### Claim 1

A system comprising:
at least one processor; and
one or more computer storage media storing computer-readable instructions thereon that when executed by the at least one processor cause the at least one processor to perform operations comprising:
accessing, from a shared memory, a dataset comprising a binary matrix of user identifiers, item identifiers, and indications identifying whether users corresponding to the user identifiers have interacted with items corresponding to the item identifiers; and
training a model by parallel processing, using a plurality of central processing units (CPUs) with access to the shared memory, each column of a design matrix determined from the binary matrix, each CPU processing one or more threads by performing a least squares analysis for each column of the design matrix within each thread, the trained model configured to provide an item recommendation from an input item interaction based on the training.

### Claim 2 (depends on 1)

The system of claim 1, wherein the least squares analysis is performed by:
determining, in each iteration, a coordinate that maximizes a gradient vector; and
updating only the coordinate corresponding to the maximized gradient vector.

### Claim 3 (depends on 1)

The system of claim 1, wherein the design matrix is determined from the product of the binary matrix and a transposed binary matrix.

### Claim 4 (depends on 1)

The system of claim 1, wherein a set of coefficients is determined during the training, and wherein the item recommendation is determined from a sum of coefficients for item identifiers corresponding to items with which there has been an interaction.

### Claim 5 (depends on 4)

The system of claim 4, wherein the coefficients determine an item score from the sum, and the item recommendation is provided based on the item identifier having the highest item score.

### Claim 6 (depends on 1)

The system of claim 1, further comprising identifying a popularity of an item within historical item data based on a number of item interactions over a time period, wherein the item recommendation is based further on the popularity.

### Claim 7 (depends on 6)

The system of claim 6, further comprising adjusting weights within the trained model using an inverse of the popularity, wherein the item recommendation is determined from the item interaction using the trained model having the adjusted weights.

### Claim 8

A method performed by one or more processors, the method comprising:
accessing, from a shared memory, a dataset comprising a binary matrix of user identifiers, item identifiers, and indications identifying whether users corresponding to the user identifiers have interacted with items corresponding to the item identifiers;
training a model comprised by parallel processing each column of a design matrix determined from the binary matrix as a separate thread by one or more central processing units (CPUs) with access to the shared memory, wherein the parallel processing processes a first column within a first thread using a least squares analysis simultaneously with a second column within a second thread using the least squares analysis; and
determining a vector comprising a set of coefficients determined from weights of the trained model, wherein the coefficients of the vector determine an item recommendation from an item interaction.

### Claim 9 (depends on 8)

The method of claim 8, wherein the least squares analysis is performed by:
determining, in each iteration, a coordinate that maximizes a gradient vector; and
updating only the coordinate corresponding to the maximized gradient vector.

### Claim 10 (depends on 8)

The method of claim 8, wherein the design matrix is determined from the product of the binary matrix and a transposed binary matrix.

### Claim 11 (depends on 8)

The method of claim 8, wherein the item recommendation is determined from a sum of the coefficients of the vector for item identifiers corresponding to items with which there has been an interaction.

### Claim 12 (depends on 11)

The method of claim 11, wherein the coefficients of the vector determine an item score from the sum, and the item recommendation is provided based on the item identifier having the highest item score.

### Claim 13 (depends on 8)

The method of claim 8, further comprising identifying a popularity of an item within historical item data based on a number of item interactions over a time period, wherein the item recommendation is based further on the popularity.

### Claim 14 (depends on 13)

The method of claim 13, further comprising adjusting weights within the trained model using an inverse of the popularity, wherein the coefficients applied to the item interaction include the adjusted weights.

### Claim 15

One or more computer storage media storing computer-readable instructions thereon that, when executed by a processor, cause the processor to perform a method comprising:
accessing, from a shared memory, a dataset comprising a binary matrix of user identifiers, item identifiers, and indications identifying whether users corresponding to the user identifiers have interacted with items corresponding to the item identifiers; and
training a model by parallel processing, using a plurality of central processing units (CPUs) with access to the shared memory, each column of a design matrix determined from the binary matrix, each CPU processing one or more threads by performing a least squares analysis for each column of the design matrix within each thread, the trained model configured to provide an item recommendation from an input item interaction based on the training.

### Claim 16 (depends on 15)

The method of claim 15, wherein the least squares analysis is performed by:
determining, in each iteration, a coordinate that maximizes a gradient vector; and
updating only the coordinate corresponding to the maximized gradient vector.

### Claim 17 (depends on 15)

The method of claim 15, wherein a set of coefficients is determined during the training, and wherein the item recommendation is determined from a sum of coefficients for item identifiers corresponding to items with which there has been an interaction.

### Claim 18 (depends on 17)

The method of claim 17, wherein the coefficients determine an item score from the sum, and the item recommendation is provided based on the item identifier having the highest item score.

### Claim 19 (depends on 15)

The method of claim 15, further comprising identifying a popularity of an item within historical item data based on a number of item interactions over a time period, wherein the item recommendation is based further on the popularity.

### Claim 20 (depends on 19)

The method of claim 19, further comprising adjusting weights within the trained model using an inverse of the popularity, wherein the item recommendation is determined from the item interaction using the trained model having the adjusted weights.
