On-the-fly Label Embedding Generation via ML Models
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing systems face challenges in efficiently selecting labels from a large or fast-changing set of potential labels, such as URLs, due to the need to store and update numerous label embeddings, which is computationally costly and resource-intensive, especially when accommodating new labels and providing personalized recommendations.
Innovation Solution
A machine-learned label embedding model generates label embeddings on the fly using feature-based data, allowing for real-time label selection without the need for constant storage and updates, and uses a Sorted-String Table to map labels to their features, while incorporating randomly sampled negative examples for improved training accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If label embeddings are stored and updated in existing systems, then label selection can be performed, but computing resources and storage burden increase significantly
Solution Approach 1:
The patent extracts the embedding generation process from pre-computed stored embeddings and performs it dynamically during inference. Instead of storing large numbers of pre-computed label embeddings, the system generates embeddings on-demand using a machine-learned model that processes label feature data, thereby reducing storage burden while maintaining label selection capability
Solution Approach 2:
The patent uses a machine-learned model to generate label embeddings by processing label feature data, effectively creating embeddings through computation rather than storage. This copying approach generates embeddings dynamically when needed, eliminating the need to store and update large numbers of pre-computed embeddings
2Measurement precision
If a large set of negative samples is used for training, then training accuracy improves, but computational cost increases
Solution Approach 1:
The patent changes the parameter of negative sample selection from exhaustive use of all available samples to random sampling. This parameter change allows the system to use a larger effective set of negative samples for training accuracy while controlling computational cost through sampling, rather than processing every possible negative sample
Data Source
AI summary
The present disclosure is directed to systems and methods that include a machine-learned label embedding model that generates feature-based label embeddings for labels in real-time, in furtherance, for example, of selection of labels relative to a particular entity. In particular, one example computing system includes both a machine-learned entity embedding model configured to receive and process entity feature data descriptive of an entity to generate an entity embedding for the entity and a machine-learned label embedding model configured to receive and process first label feature data associated with a first label to generate a first label embedding for the first label.