Live Video Preview Selection Using ML Classification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Social media platforms face challenges in helping users find interesting live video broadcasts amidst a vast number of streams, as random selection of broadcast segments often yields uninformative content due to noise such as selfies and other frequent, uninteresting segments.
Innovation Solution
The system classifies video broadcasts into relevant classes using supervised and unsupervised machine learning models, calculates the frequency of these classes across all broadcasts, and generates snippets by selecting segments with high confidence scores from classes that are less common and likely to be of interest, promoting them over frequent noise classes.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If random segments of live video broadcasts are selected for preview, then any segment can be quickly generated, but the preview content is frequently unrepresentative and uninformative due to noise segments
Solution Approach 1:
The patent replaces random mechanical selection with an automated classification system using machine learning models. The system automatically analyzes video segments, classifies them by content type, and selects previews based on classification scores rather than random chance, thereby maintaining speed while improving representativeness.
Solution Approach 2:
The system performs self-service by automatically classifying and selecting preview segments without manual intervention. The machine learning models autonomously evaluate segments and determine which ones best represent the broadcast content, eliminating the need for human curators while ensuring quality selection.
2Measurement precision
If machine learning classification is applied to select preview segments, then preview accuracy and representativeness improve, but system complexity increases
Solution Approach 1:
The patent divides the complex task of preview selection into smaller manageable components: video segmentation into discrete segments, classification of each segment into content categories, scoring based on class frequencies, and final selection based on aggregate scores. This modular approach reduces overall system complexity while maintaining high accuracy.
Solution Approach 2:
The system performs preliminary classification and scoring of video segments before final preview selection. By pre-processing segments with machine learning models and calculating class frequencies in advance, the system prepares data structures that simplify the final selection process and reduce real-time computational complexity.
3Loss of information
If frequent noise classes like selfies are excluded from previews, then preview quality improves, but the quantity of selectable segments decreases
Solution Approach 1:
The patent dynamically adjusts selection parameters based on class frequency analysis. By calculating the frequency of each content class across all broadcasts and setting thresholds for noise classes, the system adaptively determines which segments to include or exclude, maintaining quality while preserving sufficient quantity of preview segments.
Solution Approach 2:
The system applies partial exclusion by not completely removing all segments from noise classes but rather weighting them lower in the selection process. Segments from common classes like selfies are still considered but with reduced priority, allowing some flexibility while maintaining overall preview quality.
Data Source
AI summary
Systems and methods select and provide video snippets in a matrix interface. Example methods include obtaining a portion of a live video stream, assigning the portion to a class, determining that the portion is preview-eligible, and generating a snippet of the live video stream using the portion. Other example methods include determining, using a trained video classifier, a set of live broadcasts that have at least one portion that is classified as preview-eligible, calculating, for each of the live broadcasts in the set, an aggregate score for the live broadcast, selecting live broadcasts with highest aggregate scores, generating a snippet for each of the selected live broadcasts, and providing the snippets in a user interface. Methods may also include selecting portions of live video streams that are responsive to a user-provided parameter and generating a snippet for each live video stream that has a preview-eligible portion.