Recommendation Model Training With Impact-Weighted Samples
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing recommendation systems are affected by biased historical data, leading to inaccurate predictions and a Matthew effect, which negatively impacts user experience and system profitability.
Innovation Solution
A recommendation model training method that evaluates the impact of training samples on verification loss to determine weights, adjusting the model to fit data distributions more accurately and reduce biases, using a sigmoid function for convex optimization.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If the recommendation model is trained based on historical data of users, then the model can learn user preferences and improve prediction accuracy, but the historical data may contain bias leading to data distribution inconsistency and reduced prediction accuracy
Solution Approach 1:
The patent applies preliminary action by calculating impact function values for each training sample before the actual model training process. This preliminary evaluation of sample importance allows the system to pre-determine weights that will be used during training, preventing the propagation of biased data effects before they can corrupt the model learning process
Solution Approach 2:
The patent implements parameter changes by dynamically adjusting the weights of training samples based on their impact function values. By changing the weight parameter for each sample according to its measured impact on verification loss, the system transforms the uniform treatment of all training data into a differentiated approach that accounts for data quality and distribution consistency
2Ease of manufacture
If uniform weights are applied to all training samples, then the training process is simple, but biased samples can negatively affect the model and reduce prediction accuracy
Solution Approach 1:
The patent applies local quality by assigning different weights to different training samples based on their individual impact function values. Instead of applying a uniform weight across all samples, the system evaluates each sample's local contribution to verification loss and adjusts its weight accordingly, allowing important samples to have greater influence while reducing the impact of biased or less relevant samples
Solution Approach 2:
The patent implements feedback by using verification loss calculated from second training samples to inform the weighting of first training samples. The impact function value provides a feedback mechanism that measures how much each training sample affects verification performance, and this feedback is used to adjust sample weights to optimize both accuracy and distribution consistency
Data Source
AI summary
A training method includes: obtaining a first recommendation model, where a model parameter of the first recommendation model is obtained through training based on n first training samples; determining an impact function value of each first training sample with respect to a verification loss of m second training samples in the first recommendation model; determining, based on the impact function value of each first training sample with respect to the verification loss, a weight corresponding to each first training sample; and training the first recommendation model based on the n first training samples and the weights corresponding to the n first training samples, to obtain a target recommendation model.