Content Recommendation Transformers With Recursive Memory

Solution Overview

Problem

Content recommendation systems face limitations due to device constraints that restrict the amount of user content preference information that can be maintained, affecting the efficacy and accuracy of recommendations.

Innovation Solution

A recursive embedding learning platform uses machine learning models to generate data structures summarizing user preferences over time, updating these structures without increasing storage requirements, thereby maintaining an up-to-date record of user preferences.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If device storage capacity is increased to maintain more user content preference information, then recommendation accuracy is improved, but device complexity and cost increase

Engineering Contradiction:

Improverecommendation accuracyVSAvoiddevice complexity

Core Design Contradiction:

Measurement precisionVSDevice complexity

Solution Approach 1:

The patent extracts and stores only the essential user preference information in a compressed format within the data structure, separating the critical preference data from the complete viewing history. This allows the device to maintain accurate recommendations without storing all historical data, thus reducing storage requirements and device complexity while preserving recommendation accuracy.

Inventive Principle:

Principle #2Taking out (Extraction)

Solution Approach 2:

The patent creates a simplified copy or representation of user preference information within the data structure, which captures the essential patterns and preferences without replicating the complete viewing history. This copy allows the system to make accurate recommendations based on condensed preference data rather than requiring full historical records.

Inventive Principle:

Principle #26Copying

2Measurement precision

If the amount of user content preference information maintained is increased, then recommendation accuracy is improved, but storage requirements increase

Engineering Contradiction:

Improverecommendation accuracyVSAvoidstorage requirements

Core Design Contradiction:

Measurement precisionVSQuantity of substance

Solution Approach 1:

The patent extracts and stores only the essential user preference information in a compressed format within the data structure, separating the critical preference data from the complete viewing history. This allows the device to maintain accurate recommendations without storing all historical data, thus reducing storage requirements and device complexity while preserving recommendation accuracy.

Inventive Principle:

Principle #2Taking out (Extraction)

Solution Approach 2:

The patent transforms the storage representation of user preference information by changing the data parameters and format. The system stores preferences in a compressed, structured format that reduces the quantity of data while maintaining the essential information needed for accurate recommendations, effectively changing how the data is represented and stored.

Inventive Principle:

Principle #35Parameter changes

3Measurement precision

If complete viewing history is stored for accurate recommendations, then recommendation accuracy is improved, but loss of information is reduced

Engineering Contradiction:

Improverecommendation accuracyVSAvoidinformation loss

Core Design Contradiction:

Measurement precisionVSLoss of information

Solution Approach 1:

The patent extracts and stores only the essential user preference information in a compressed format within the data structure, separating the critical preference data from the complete viewing history. This allows the device to maintain accurate recommendations without storing all historical data, thus reducing storage requirements and device complexity while preserving recommendation accuracy.

Inventive Principle:

Principle #2Taking out (Extraction)

Data Source

PatentUS20250307608A1Systems And Methods For Providing Content Recommendations

Publication Date: 2025.10.02 COMCAST CABLE COMM LLC

AI summary

Systems, apparatuses, and methods are described for providing content recommendations using recursive learning transformers. A computing platform may train machine learning models (e.g., transformers) to generate content recommendations based on data structures summarizing user content preference information. The machine learning models may utilize a long-term memory data structure comprising a summary of user content preference information over any period of time (e.g., the entire length of time a user is associated with the computing platform) to generate the content recommendations. The long-term memory data structure may be recursively updated to maintain the summary without increasing storage requirements.