Video Embedding Clustering for Efficient Third-Party Matching
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current video distribution platforms face challenges in efficiently matching and recommending third-party video items to users due to lack of viewing history, leading to difficulties in determining their relevance.
Innovation Solution
A system and method that generates embeddings for both first-party and third-party video items using neural networks, groups third-party items into clusters based on these embeddings, and matches first-party items to clusters to reduce comparison complexity, thereby improving matching speed and resource efficiency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If all third-party video items are individually compared with first-party video items, then matching accuracy is improved, but computing time and resource consumption increase significantly
Solution Approach 1:
The patent segments the large set of third-party video items into multiple clusters based on embedding similarity. Instead of comparing each third-party item individually with all first-party items, the system first creates compact representations (cluster centers) of groups of similar items. This segmentation reduces the search space while preserving matching accuracy through the hierarchical comparison approach.
Solution Approach 2:
The patent introduces cluster centers as intermediary representations between individual video items and the matching process. These cluster centers serve as mediators that capture the essential characteristics of groups of video items, enabling efficient preliminary filtering before detailed item-to-item comparison. This intermediary layer significantly reduces computing time while maintaining accuracy.
2Measurement precision
If all third-party video items are individually compared with first-party video items, then matching accuracy is improved, but device complexity and resource requirements increase
Solution Approach 1:
The system segments the complex matching task into two simpler sub-tasks: (1) clustering third-party items based on their embeddings, and (2) comparing cluster centers with first-party items followed by detailed comparison only for promising matches. This segmentation reduces system complexity by breaking down the intractable all-pairs comparison into manageable stages.
Solution Approach 2:
The patent performs preliminary clustering of third-party video items before the actual matching process. This preliminary action organizes the data into structured groups with representative centers, simplifying the subsequent matching operation. By pre-processing the third-party items into clusters, the system reduces the complexity of the main matching task.
3Measurement precision
If viewing history factors are used to determine video relevance, then recommendation accuracy is improved, but applicability to new third-party video items deteriorates
Solution Approach 1:
The patent generates embeddings for third-party video items that replicate or copy the representation format used for first-party items with viewing history. This embedding-based copying allows the system to apply the same matching methodology to both types of items, enabling third-party items without viewing history to be processed using identical techniques, thus improving adaptability while maintaining accuracy.
Solution Approach 2:
The patent creates a universal embedding-based matching system that works for both first-party video items (with viewing history) and third-party video items (without viewing history). The embedding representation serves multiple functions: it captures content semantics, enables similarity computation, and works uniformly across different item types and data availability scenarios, achieving both accuracy and versatility.
Data Source
AI summary
Techniques for efficiently matching two sets of video items are provided. In on technique, an embedding is generated for each video item in each set. For the first set of video items, multiple groups are generated. The first set of video items may have a relatively little amount of metadata information for them. Each video item in the first set is assigned to one of the groups. Then, for each video item in the second set, one of the groups is selected based on embedding similarity. For each video item in the selected group, an embedding similarity is determined between that video item in the selected group and the video item in the second set. If the embedding similarity is above a certain threshold, then an association is generated for that pair of video items.