Recommendation Model Pruning for Low-Frequency Feature Lookup
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing recommendation models suffer from high memory occupancy and inefficiency due to the inclusion of numerous high-dimensional sparse features with low impact on click-through rates, leading to prolonged lookup times and reduced operational efficiency.
Innovation Solution
The proposed method involves removing model parameters corresponding to features with frequencies less than or equal to a threshold from the recommendation model, retaining only those with higher frequencies, thereby reducing model size and memory occupancy while maintaining accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If model parameters corresponding to features with frequencies less than or equal to a threshold are retained in the recommendation model, then the model comprehensively covers all feature dimensions, but memory occupation increases and lookup time increases
Solution Approach 1:
The patent extracts and removes model parameters corresponding to features with frequencies less than or equal to a threshold from the recommendation model. This extraction process eliminates unnecessary parameters that contribute to memory occupation and lookup time without significantly impacting prediction accuracy, as these low-frequency features have minimal influence on user click behavior.
Solution Approach 2:
The patent changes the parameter set of the recommendation model by selectively removing parameters associated with low-frequency features. This parameter change optimizes the model by reducing its size and complexity while maintaining the essential predictive capabilities through the retention of high-frequency feature parameters that have greater impact on click-through rate prediction.
2Productivity
If model parameters corresponding to features with frequencies less than or equal to a threshold are removed from the recommendation model, then memory occupation and lookup time are reduced, but the model may lose some feature coverage
Solution Approach 1:
The patent extracts and removes model parameters corresponding to features with frequencies less than or equal to a threshold from the recommendation model. This extraction process eliminates unnecessary parameters that contribute to memory occupation and lookup time without significantly impacting prediction accuracy, as these low-frequency features have minimal influence on user click behavior.
Solution Approach 2:
The patent changes the parameter set of the recommendation model by selectively removing parameters associated with low-frequency features. This parameter change optimizes the model by reducing its size and complexity while maintaining the essential predictive capabilities through the retention of high-frequency feature parameters that have greater impact on click-through rate prediction.
3Adaptability or versatility
If all model parameters are retained for comprehensive feature coverage, then the model can handle diverse feature dimensions, but operational efficiency decreases due to slow lookup times
Solution Approach 1:
The patent extracts and removes model parameters corresponding to features with frequencies less than or equal to a threshold from the recommendation model. This extraction process eliminates unnecessary parameters that contribute to memory occupation and lookup time without significantly impacting prediction accuracy, as these low-frequency features have minimal influence on user click behavior.
Solution Approach 2:
The patent changes the parameter set of the recommendation model by selectively removing parameters associated with low-frequency features. This parameter change optimizes the model by reducing its size and complexity while maintaining the essential predictive capabilities through the retention of high-frequency feature parameters that have greater impact on click-through rate prediction.
Data Source
AI summary
Provided is a method for recommending information. The method includes acquiring a click through rate of recommended information by predicting a feature vector of the recommended information by a recommendation model, wherein model parameters corresponding to features that are the same as features of the feature vector and have frequencies less than or equal to a first threshold are removed from the recommendation model; and issuing the recommended information based on the click through rate.