LLM Resource Screening for Cold-Start Recommendation Accuracy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing resource recommendation systems lack accuracy in selecting resources of interest to users, particularly for new or less-known resources due to insufficient historical data and user feedback.
Innovation Solution
A resource screening method utilizing a large language model trained through machine learning to generate interest description information based on historical behavior and model indication, enabling the selection of resources with higher similarity scores to user interests.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional resource recommendation systems are used, then the system structure is simple, but the accuracy of resource selection is poor especially for new or less-known resources
Solution Approach 1:
A large language model is introduced as an intermediary component between the resource information and user interest representation. The LLM processes historical behavior information and generates interest description information that serves as a mediator to bridge the gap between available resources and user preferences, enabling accurate recommendation even for new resources without requiring complex traditional recommendation systems.
Solution Approach 2:
The system changes the parameter representation from traditional user-profile-based vectors to natural language interest description information generated by the LLM. This parameter transformation allows the system to capture user interests more accurately through semantic understanding of historical behaviors, improving recommendation accuracy without proportionally increasing system complexity.
2Reliability
If traditional recommendation systems are used, then the system is easier to implement, but it cannot accurately recommend new or less-known resources due to insufficient historical data
Solution Approach 1:
The system performs preliminary action by using the large language model to generate interest description information from available historical behavior data before the actual recommendation is made. This preliminary processing transforms sparse or insufficient historical data into comprehensive interest representations, enabling reliable recommendations for new resources that would otherwise be impossible to recommend accurately.
Solution Approach 2:
The system creates a semantic copy or representation of user interests through natural language descriptions generated by the LLM. Instead of directly using raw historical data which may be insufficient, the system creates an enriched semantic representation that captures the essence of user preferences, allowing accurate matching with new resources even when direct historical data is limited.
3Measurement precision
If more historical data is collected to improve recommendation accuracy, then the accuracy improves, but the time required and data processing complexity increases
Solution Approach 1:
The system replaces traditional mechanical data processing approaches with a large language model-based semantic processing system. Instead of manually analyzing and processing large volumes of historical data through complex algorithms, the LLM automatically transforms raw historical behavior information into accurate interest descriptions through its inherent semantic understanding capabilities, significantly reducing processing time while maintaining or improving accuracy.
Data Source
AI summary
A resource screening method includes obtaining historical behavior information of an object and model indication information, the historical behavior information indicating that the object has interacted with a resource, the model indication information indicating that a large language model generates interest description information based on input information, and the interest description information describing a resource of interest to the object; prompt information indicating a resource with which the object has interacted is generated based on the historical behavior information; the interest description information is generated based on the model indication information and the prompt information by using the large language model; a first resource is selected from among resources based on the interest description information and information about the resources, wherein a similarity score between the first resource and the interest description information is greater than that of a second resource.