Content Embeddings for Video Recommendation Relevance
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing video recommendation systems often provide irrelevant recommendations due to skewed collaborative filtering results, especially when popular videos are watched together, leading to a lack of relevance in user recommendations.
Innovation Solution
A recommendation system that generates content-level recommendations by processing text and tags metadata to create content embeddings, which represent videos in a vector space, allowing for more accurate similarity-based recommendations independent of user watch behavior.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If collaborative filtering is used to generate recommendations based on user watch behavior, then the system can provide personalized recommendations, but the recommendations become skewed when popular videos are watched together, reducing relevance
Solution Approach 1:
The patent introduces content embeddings as an intermediary representation that mediates between user behavior data and recommendation generation. Instead of directly using collaborative filtering on raw user-watch data, the system transforms video content into embedding vectors that capture intrinsic content similarities, serving as a mediator that decouples recommendations from popularity-based skewing while preserving personalization capabilities
Solution Approach 2:
The patent replaces the mechanical collaborative filtering mechanism (which relies on user behavior patterns) with a content-based embedding system. By substituting the user-behavior-dependent collaborative filtering approach with content-embedding-based similarity measurement, the system eliminates the harmful dependency on popular video co-viewing patterns while maintaining recommendation personalization
2Quantity of substance
If collaborative filtering relies on popular videos being watched together, then data availability is improved, but the similarity measurement becomes inaccurate due to popularity skewing
Solution Approach 1:
The patent extracts the essential content features from videos and represents them as embeddings, separating the content similarity signal from the noisy user behavior data. By taking out only the relevant content characteristics and representing them in embedding space, the system achieves accurate similarity measurement without being influenced by the quantity of user watch data or popularity patterns
Solution Approach 2:
The patent transforms the similarity measurement parameter from user-behavior-based co-viewing frequency to content-embedding-based vector similarity. By changing the fundamental parameter used for similarity assessment from population-level watch patterns to content-level embedding distances, the system maintains measurement accuracy independent of data quantity or video popularity
Data Source
AI summary
In some embodiments, a method maps attributes of metadata for a plurality of content instances to metadata nodes. The metadata nodes are connected to type nodes that define a type of metadata for each metadata node and a respective content node for a respective content instance. The method generates a plurality of sample paths using the content nodes, the metadata nodes, and the type nodes from the mapping of attributes of the metadata. A similarity of content nodes is analyzed using the plurality of sample paths. Then, the method generates a representation of each of the plurality of content instances using the similarity of the content nodes. The representation represents the similarity between content instances in the plurality of content instances.