Machine-Learned Models for Real-Time Content Throttling
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current content throttling techniques are inefficient as they do not consider machine-learned models when determining which additional content items to throttle, leading to suboptimal content selection and resource utilization.
Innovation Solution
The system uses machine-learned models to generate user action predictions and relevance scores, organizing them into distributions to select a threshold for content delivery groups, allowing for real-time content item selection based on resource utilization and relevance, thereby improving content selection efficiency and user experience.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If traditional content throttling techniques are used without machine-learned models, then the system is simpler to implement, but content selection efficiency and resource utilization deteriorate
Solution Approach 1:
The system performs preliminary actions by pre-calculating and storing user action predictions and relevance scores in machine-learned models before content selection events occur. These pre-computed values are organized into distributions and thresholds are established in advance, allowing the throttling system to make rapid decisions during real-time content selection without performing complex calculations at the moment of selection, thereby improving efficiency while managing complexity through advance preparation
Solution Approach 2:
The patent introduces machine-learned models as intermediary components between the content selection system and the throttling logic. These models serve as mediators that pre-process user behavior data and content characteristics to generate predictions and relevance scores, which then guide the throttling decisions. This intermediary layer handles the computational complexity separately, allowing the core content selection system to benefit from improved efficiency without directly managing the complexity of predictive analytics
2Measurement precision
If content throttling is based solely on resource allocation without considering machine-learned models, then the processing requirements are lower, but content selection quality and user experience deteriorate
Solution Approach 1:
The system pre-computes user action predictions and relevance scores using machine-learned models before content selection events. These predictions are organized into distributions and thresholds are established in advance based on historical data and resource constraints. During real-time content selection, the system only needs to compare current predictions against pre-established thresholds, significantly reducing processing requirements while maintaining high content selection quality through the use of sophisticated predictive models that were trained offline
Solution Approach 2:
The patent implements partial throttling by applying machine-learned models selectively to determine which content items should be throttled based on predicted user actions and relevance scores. Rather than processing all content items uniformly, the system focuses computational resources on evaluating and throttling only those items where the machine-learned models indicate potential resource over-utilization or low expected value, thereby achieving high selection quality while minimizing overall processing resource consumption
3Productivity
If all allocated content item resources are utilized without throttling, then short-term content visibility is maximized, but long-term resource management and system performance deteriorate
Solution Approach 1:
The patent implements a feedback mechanism where machine-learned models continuously predict user actions for content items and these predictions are fed back into the throttling system. The system monitors resource utilization patterns and uses the predicted user actions to dynamically adjust throttling decisions. This feedback loop allows the system to identify when content items are consuming resources without generating proportional user engagement, enabling proactive throttling decisions that optimize long-term resource utilization efficiency while preventing resource exhaustion and maintaining sustainable content item lifecycle management
Data Source
AI summary
Techniques for using machine-learned models to throttle content are provided. In one technique, based on multiple selection events, a distribution of relevance measures is computed, where the relevance measures are associated with the content item selection events. The relevance measures may be generated by one or more machine-learned models. Based on the computed distribution, a threshold relevance measure is computed. Thereafter, a request for content is received over a computer network. In response, a computer system performs, in real-time, multiple steps. For example, an identity of an entity that is associated with the request is identified and, based on that identity, multiple content delivery groups are identified. A relevance measure of one of the content delivery groups relative to the entity is determined and compared to the threshold relevance measure. The content delivery group is selected only after determining that the relevance measure is above the threshold relevance measure.