Offline Embedding Retrieval for Accurate B2B Content Targeting
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing advertising systems face challenges in accurately identifying target user groups due to reliance on limited user information, inflexibility in product categories, and inefficiency in adapting to emerging trends, particularly in business-to-business (B2B) settings.
Innovation Solution
A target content item group generation system that uses a two-tower model to generate member and content item embeddings, leveraging member activity, outreach, and information data to calculate similarity scores, allowing for dynamic content item embeddings and accurate prediction of stakeholder interest in B2B environments.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional advertising systems use limited user information for targeting, then device complexity is reduced, but measurement precision of target audience identification deteriorates
Solution Approach 1:
The patent segments user information into multiple dimensions including user activities, user attributes, and contextual information. Each dimension is processed separately through embedding models to create comprehensive user profiles, enabling precise target identification without overwhelming system complexity through structured data organization
Solution Approach 2:
The patent introduces embedding models as intermediary components that transform raw user information into standardized vector representations. These embeddings serve as mediators between diverse data sources and the matching algorithm, enabling accurate comparisons while maintaining system modularity and manageability
2Adaptability or versatility
If conventional systems use fixed product categories, then device complexity is reduced, but adaptability to emerging trends deteriorates
Solution Approach 1:
The patent implements dynamic product category structures where category embeddings are generated on-demand based on content item characteristics rather than relying on fixed predetermined categories. This allows the system to adapt to emerging trends and novel product types while maintaining computational efficiency through embedding caching and similarity-based category inference
Solution Approach 2:
The patent changes the parameter representation from discrete fixed category labels to continuous embedding vectors. This transformation enables gradient-based optimization and flexible category boundaries, allowing the system to adapt to new trends by learning new category representations from data without requiring structural system changes
3Productivity
If conventional systems process large amounts of data in real-time, then productivity is improved, but use of energy increases
Solution Approach 1:
The patent performs preliminary actions by pre-computing and storing user embeddings and content embeddings in advance. During real-time advertising delivery, the system only needs to retrieve pre-computed embeddings and perform simple similarity calculations, dramatically reducing computational resource requirements while maintaining high target identification efficiency
Solution Approach 2:
The patent creates compact embedding vector copies of user profiles and content items that capture essential characteristics in a space-efficient format. These embedding copies enable rapid similarity computations with minimal computational resources compared to processing full raw data, achieving high productivity with reduced energy consumption
4Measurement precision
If conventional systems use simple matching algorithms, then device complexity is reduced, but measurement precision of user-content matching deteriorates
Solution Approach 1:
The patent replaces traditional mechanical rule-based matching algorithms with embedding-based similarity computations. This substitution leverages vector space geometry and distance metrics to capture complex user-content relationships, achieving superior matching precision while maintaining computational simplicity through efficient dot product or cosine similarity calculations
Data Source
AI summary
The present disclosure relates to systems, non-transitory computer-readable media, and methods for generating a content item group comprising members that have interest in a content item. In particular, the disclosed systems can generate a member embedding by leveraging member activity feature data and member information feature data. The disclosed systems can further generate a content item embedding reflecting content item feature data. The disclosed systems may generate a similarity score between the member embedding and the content item embedding. Based on the similarity score meeting a threshold similarity score, the disclosed system can determine to include a member within a target content item group.