Multi-Stage Recall Model for Resource-Efficient Data Collection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional machine learning models for content recommendation are resource-intensive and struggle to efficiently select top-K items from large datasets within practical latency and computational constraints, especially in applications like social networking and advertising, where they must make millions of determinations quickly.
Innovation Solution
A multi-stage recall optimization model is employed, comprising a less complex early stage model and a more complex target model, where the early stage model predicts the likelihood of items being in the top-K based on the target model's scores, using soft labeling and random sampling to reduce computational burden.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional machine learning models are used for content recommendation, then accuracy can be maintained, but computational cost and resource consumption increase significantly
Solution Approach 1:
The patent segments the recommendation system into multiple stages: an early stage model that performs initial filtering and ranking of candidate items, and a target model that processes only the top-ranked items from the early stage. This segmentation allows the computationally expensive target model to operate on a reduced dataset while maintaining overall recommendation accuracy, thereby reducing total computational cost.
Solution Approach 2:
The early stage model performs a preliminary ranking of all candidate items, and only the top-K items (where K is a small fraction of total candidates) are passed to the target model. This partial action approach ensures that the expensive target model processes only a subset of items, significantly reducing computational resources while preserving accuracy for the most relevant recommendations.
2Reliability
If conventional machine learning models process large datasets to select top-K items, then completeness of evaluation is improved, but processing time and latency increase
Solution Approach 1:
The patent divides the item selection process into two segments: the early stage model quickly ranks all candidate items to identify potential top-K candidates, and the target model then evaluates only these candidates in detail. This segmentation enables the system to maintain reliable evaluation of top-K items while drastically reducing processing time compared to evaluating all candidates with the target model.
Solution Approach 2:
The early stage model performs preliminary ranking and filtering of candidate items before they reach the target model. This preliminary action pre-sorts and pre-filters the dataset, so that when the target model processes items, it only needs to evaluate a small subset that has already been identified as potential top-K candidates, reducing overall processing time while maintaining evaluation completeness.
3Measurement precision
If the target model is used to score all candidate items, then ranking accuracy is improved, but computational complexity increases
Solution Approach 1:
The patent segments the ranking process into an early stage model that performs initial scoring and filtering, and a target model that performs detailed scoring only on the filtered subset. This segmentation allows the system to maintain high ranking accuracy through the target model while reducing computational complexity by limiting the target model's input to only the most promising candidates identified by the early stage.
Solution Approach 2:
The early stage model acts as an intermediary between the large set of candidate items and the computationally expensive target model. It pre-processes and filters candidates, transforming the original large dataset into a manageable subset that the target model can process efficiently, thereby reducing computational complexity while preserving ranking accuracy for the final recommendations.
4Quantity of substance
If millions of candidate items are evaluated to make rapid recommendations, then coverage of candidates is improved, but processing speed decreases
Solution Approach 1:
The patent segments the candidate evaluation process into broad coverage at the early stage (evaluating all candidates through a lightweight model) followed by detailed evaluation of a small subset at the target stage. This segmentation enables the system to maintain coverage of all candidate items while achieving high processing speed, as the expensive detailed evaluation is performed only on the final top-K selections rather than all candidates.
Solution Approach 2:
The system performs complete evaluation (excessive action) at the early stage using a lightweight model to ensure all candidates are considered, then performs partial evaluation only on the top-ranked candidates using the target model. This approach maintains coverage of all candidates while achieving high processing speed by limiting the computationally intensive operations to a small subset.
Data Source
AI summary
Systems, methods, and non-transitory computer-readable media can receive a set of candidate training items for training an early stage model in a multi-stage recall optimization model, wherein the multi-stage recall optimization model comprises the early stage model and a target model. A random subset of the candidate training items is selected from the set of candidate training items. For each training item in the subset of candidate training items, a score is determined based on the target model. Each training item in the subset of candidate training items is labeled with a label based on a probability of the training item being a top-K of the set of candidate training items had the set of candidate training items been scored based on the target model.