User Embeddings for Capturing Interest Relationships
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Traditional user profile representations fail to capture complex relationships among different interests, limiting their effectiveness in personalization for online services.
Innovation Solution
The method involves computing user embeddings that capture both semantic and temporal relationships among user interests, using semantic embeddings derived from textual information and temporal embeddings from co-occurring content within a specified period, to provide personalized content.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of information
If traditional user profile representations (list with scores) are used, then the system is simple to implement, but it fails to capture complex relationships among different interests
Solution Approach 1:
The patent transforms user interests from a discrete list representation to continuous embedding vectors, changing the parameter space from categorical scores to dense numerical representations. This allows the system to capture complex relationships through vector operations while maintaining computational efficiency through optimized embedding algorithms.
Solution Approach 2:
The patent introduces temporal dimension to user interest representation by computing embeddings that incorporate both current interests and temporally related interests. This dimensional expansion from static to dynamic representation enables capturing evolving user preferences and relationships over time.
2Loss of information
If user profiles include only declared and estimated interests, then the profile construction is straightforward, but it cannot capture temporally related interests
Solution Approach 1:
The patent pre-computes and stores embeddings for various temporal windows (e.g., 7-day, 30-day, 90-day periods) in advance. When generating user interest embeddings, the system retrieves these pre-computed temporal embeddings and combines them with current interest embeddings, avoiding real-time complex temporal analysis while capturing temporally related interests.
3Measurement precision
If the system uses detailed user interest representations, then personalization accuracy improves, but computational resources increase
Solution Approach 1:
The patent implements a two-stage embedding generation process: first generating comprehensive user interest embeddings with full temporal context for high accuracy, then applying dimensionality reduction or selective feature extraction to create condensed representations for actual content recommendation. This partial action approach maintains precision where needed while reducing overall computational energy consumption.
Data Source
AI summary
The present teaching relates to method, system, medium, and implementations for personalized content service. Information related to a user is first obtained with a user profile indicative of multiple interests of the user. User embeddings are computed with respect to some interests of the user based on interest embeddings of such interests to capture semantics of such interests as well as additional interests temporally related to the interests. Personalized content is identified based on the user embeddings and is provided to the user.