Two Tower Network Extension for Recommendation Latency
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing recommendation systems face challenges in providing accurate and scalable recommendations for users and content items in real-time, especially in large-scale application software systems, where predicting user interest and making relevant suggestions is crucial for engagement and attendance.
Innovation Solution
An extended two-tower network architecture is introduced, comprising a content item sub-model, a user sub-model, and a fusion sub-model, which jointly predict user interest by combining user and content item embeddings with interaction features, allowing for both content item and candidate invitee recommendations with improved relevance and scalability.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a traditional recommendation system processes user and content item data in real-time, then recommendation relevance can be maintained, but system latency increases and throughput decreases
Solution Approach 1:
The patent pre-generates and stores embeddings for users and content items before they are needed for recommendations. These embeddings are created in advance using historical data and user interactions, then stored in a database. When a recommendation request arrives, the system simply retrieves pre-computed embeddings and interaction features, avoiding time-consuming real-time computations and significantly reducing latency while maintaining recommendation quality
2Adaptability or versatility
If recommendation systems process large volumes of user and content item data, then comprehensive recommendations can be provided, but system throughput decreases
Solution Approach 1:
The patent divides the recommendation system into independent sub-models: a user sub-model that processes user data, a content item sub-model that processes content data, and an interaction sub-model that combines them. Each sub-model independently generates embeddings for its respective data type. This segmentation allows parallel processing of user and content data, improving throughput while maintaining comprehensive recommendation coverage across multiple data types
Solution Approach 2:
The system pre-computes and stores embeddings for all users and content items in advance, along with pre-computed interaction features. This preliminary processing transforms complex real-time computation into simple database lookups during recommendation generation, dramatically increasing system throughput while still providing comprehensive recommendations across diverse user and content item data
Data Source
AI summary
An extended two tower network is used to make both a recommendation of a content item that a given user may be interested in, and a recommendation of a user that may be interested in a given content item. The extended two tower network includes a content item sub-tower, a user sub-tower, and a fusion sub-model that are jointly trained to predict a probability that a given user is interested in a given content item. The content item sub-tower and the user sub-tower are used to make an initial prediction that the given user will be interested in the given content item. The initial prediction is then input to the fusion sub-model to make a final prediction. In the case of a candidate invitee recommendation, the initial prediction may be combined with one or more interaction features and the combination input to the fusion sub-model to make the final prediction.