Watch-time clustering for video search relevance
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current video search techniques are inefficient, leading to unnecessary usage of network resources as users often revise queries multiple times to find relevant videos, resulting in unsatisfactory search results and wastage of processing and bandwidth resources.
Innovation Solution
A system that utilizes watch-time information and creator performance scores to rank and select video content based on user characteristics, such as demographic and behavioral data, to provide more relevant search results and recommendations, thereby reducing the need for iterative querying and conserving resources.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If video search results are selected based on topic comparison only, then the search system is simple to implement, but the search results are not relevant enough to user interest, causing users to revise queries multiple times
Solution Approach 1:
The system pre-computes watch-time statistics and creator performance scores before search queries are submitted. This preliminary action prepares ranking data in advance, allowing the system to quickly retrieve and use pre-analyzed metrics during actual search operations, thereby improving relevance without proportionally increasing real-time complexity
Solution Approach 2:
The search system segments the ranking process into multiple independent components: topic matching, watch-time-based ranking, and creator performance scoring. Each component operates separately and contributes to the final ranking, allowing the system to maintain simplicity while incorporating multiple factors for improved relevance
2Measurement precision
If users revise queries iteratively to find relevant videos, then more relevant results can be found, but network resources and processing power are wasted
Solution Approach 1:
The system incorporates feedback from user watch behavior into the search ranking process. By analyzing how long users watch videos and which creators they engage with, the system learns and adjusts rankings to better predict user interest, reducing the need for iterative query revisions and associated resource waste
Solution Approach 2:
The system pre-computes creator performance scores and watch-time metrics before searches are executed. This preliminary preparation ensures that when users submit queries, the system can immediately provide well-ranked results based on historical behavior data, preventing the need for multiple query revisions and the associated resource consumption
3Ease of operation
If expected videos are not among top results, then users can still find them through iterative searching, but valuable network resources are consumed unnecessarily
Solution Approach 1:
The system changes the ranking parameters from topic-based only to include watch-time duration and creator performance scores. This parameter change reorders results to prioritize videos that align with user demonstrated interests, ensuring expected videos appear in top results and reducing unnecessary resource consumption from iterative searching
Data Source
AI summary
This document describes, among other things, systems, methods, devices, and other techniques for using information about how long various videos were presented at client devices to determine subsequent video recommendations and search results. In some implementations, a computing can include a modeling apparatus, a front-end server, a request manager, one or more video file storage devices, a video selector, or a combination of some or all of these. The video selector can select video content for a particular digitized video among a plurality of digitized videos to serve to a computing device responsive to a request. The selection can be based at least in part on how long the particular digitized video has been presented at client devices associated with users having characteristics that match one or more characteristics of the user that submitted the request for video content, as indicated by the modeling apparatus.