On-Device User Interest Quantification via Text Embedding
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for estimating user interest face challenges in quantifying interest in items not previously defined and require significant resources for designing and maintaining multiple interest estimation models, with data transmission and resource consumption being substantial issues.
Innovation Solution
An electronic device generates embedding vectors through text embedding using user data and calculates weights based on usability information, allowing for on-device quantification of user interest in queries without the need for external servers and additional feature extraction processes.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If multiple interest estimation models are designed and maintained for different interest items, then user interest quantification accuracy is improved, but device complexity and resource consumption increase
Solution Approach 1:
The patent applies universality by designing a single interest estimation model that can handle multiple interest items through text embedding. Instead of creating separate models for each interest item (app usage, website visit, search information), the system uses one universal model that processes text data from all sources by converting them into embedding vectors, thereby reducing model management complexity while maintaining quantification accuracy across diverse interest domains
Solution Approach 2:
The patent applies parameter changes by transforming different types of user data (app usage information, website visit information, search information) into a unified parameter space through text embedding. By converting diverse data types into embedding vectors that represent semantic meaning, the system enables a single model to process varied inputs by changing the representation parameters rather than requiring separate models for each data type
2Measurement precision
If user data is transmitted to external servers for interest estimation, then quantification accuracy is improved, but data transmission overhead and processing time increase
Solution Approach 1:
The patent applies the extraction principle by removing the dependency on external servers for interest estimation. The system extracts and implements the interest estimation functionality directly within the electronic device, allowing user data to be processed locally without transmission to external servers. This extraction of the estimation function from the server environment eliminates data transmission overhead while maintaining estimation accuracy through the embedded model and vector space operations
3Adaptability or versatility
If a large number of interest estimation models are maintained for different interest items, then coverage of interest items is improved, but resource consumption for model design and maintenance increases
Solution Approach 1:
The patent applies universality by creating a single multi-functional interest estimation model that can estimate user interest across multiple interest items (apps, websites, searches) through text embedding. This universal model processes diverse user data by converting them into embedding vectors and calculating similarity in a unified vector space, thereby achieving broad interest item coverage without requiring separate models for each interest domain, significantly reducing the computational resources needed for model design, training, and maintenance
Solution Approach 2:
The patent applies merging by combining multiple interest estimation functionalities into a single unified model. Instead of maintaining separate models for app usage analysis, website visit analysis, and search information analysis, the system merges these functionalities into one model that processes all user data types through text embedding and similarity calculation in a shared vector space, reducing the overall resource consumption for model management while maintaining comprehensive interest item coverage
Data Source
AI summary
An electronic device includes a memory configured to store a plurality of types of user data, and a processor configured to be operatively connected to the memory, wherein the processor is configured to generate an embedding vector through text embedding using first text data which is included in the user data and is related to content of the user data for each type of the user data; calculate a weight for the embedding vector using information which is included in the user data and is related to usability of the user data for each type of the user data; when a query is input, generate a query vector through the text embedding using second text data included in the query; and quantify user interest in the query for each type of the user data based on the embedding vector, the weight for the embedding vector, and the query vector.