Multi-Granular Embeddings for Efficient Content Tagging
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Content distribution platforms face challenges in efficiently tagging and determining the relevance of third-party content items, as they lack viewing history data, leading to difficulties in predicting user interest and accurately recommending such content.
Innovation Solution
A system and method using machine-learned embeddings of varying granularities to associate entity names with content items, converting multi-label classification into binary classification, generating feature values based on similarity measures, and automatically creating negative training instances to improve the accuracy of content item tagging and recommendation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If multi-label classification is used to tag content items with multiple entity names, then the comprehensiveness of content tagging is improved, but the computing resources and complexity required for processing increase significantly
Solution Approach 1:
The patent segments the multi-label classification problem into multiple binary classification problems by processing entity names in batches. Instead of evaluating all possible entity names simultaneously for each content item, the system divides the entity name space into manageable groups and processes them sequentially, reducing computational complexity while maintaining comprehensive tagging capability
Solution Approach 2:
The patent applies partial action by evaluating only a subset of entity names for each content item rather than all possible entity names. The system generates candidate entity names based on content characteristics and evaluates only those relevant candidates through the machine-learned model, reducing computational overhead while maintaining sufficient tagging accuracy
2Measurement precision
If viewing history data is collected and processed to determine content relevance, then the accuracy of content recommendation is improved, but the data processing complexity and privacy concerns increase
Solution Approach 1:
The patent introduces machine-learned embeddings as an intermediary representation that captures content characteristics without requiring direct processing of raw viewing history data. The embeddings serve as a compressed, processed representation that enables relevance prediction while reducing the complexity of working with raw user behavior data
Solution Approach 2:
The patent transforms the data representation from raw viewing history logs to normalized embedding vectors. This parameter transformation consolidates complex user behavior patterns into compact numerical representations that can be efficiently processed and compared, reducing data processing complexity while maintaining recommendation accuracy
3Adaptability or versatility
If third-party content items are ingested without viewing history, then the platform's content capacity and versatility are improved, but the ability to predict user interest and recommend relevant content deteriorates
Solution Approach 1:
The patent applies preliminary action by pre-computing and storing machine-learned embeddings for content items before they are recommended to users. This allows the system to quickly evaluate relevance for third-party content without requiring real-time viewing history analysis, enabling efficient ingestion and immediate relevance assessment for new content
Solution Approach 2:
The patent creates a computational copy of content items in the form of machine-learned embeddings that capture essential characteristics. These embedding copies enable rapid relevance prediction and comparison without requiring access to or processing of the original content data, allowing efficient handling of large volumes of third-party content
Data Source
AI summary
Efficient tagging of content items using content embeddings are provided. In one technique, multiple content items are stored a content embedding for content item is stored. Entity names are also stored along with an entity name embedding for each entity name. For each content item, (1) multiple content embeddings that are associated with the content item are identified; (2) a subset of the entity names is identified; and (3) for each entity name in the subset, (i) an embedding of the entity name is identified, (ii) similarity measures are generated based on the entity name embedding and the multiple content embeddings, (iii), a distribution of the similarity measures is generated, (iv) feature values are generated based on the distribution, (v) the feature values are input into a machine-learned classifier, and (vi) based on output from the classifier, it is determined whether to associate the entity name with the content item.