Recommendation Model Training via Interpolation Bias Correction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing recommendation models face accuracy issues due to location bias and selection bias in training data, leading to reduced model training accuracy and negative impacts on user experience and revenues.
Innovation Solution
A recommendation model training method that uses an interpolation model to predict user actions by processing attribute information of users and recommended objects, with training samples obtained when objects are randomly displayed to users, reducing bias and improving model accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If training data is collected based on recommended display (non-random), then model training efficiency is improved, but training data bias increases reducing model accuracy
Solution Approach 1:
The patent introduces an interpolation model as an intermediary component between the recommendation model and training data. This interpolation model generates prediction labels for training samples by interpolating between positive samples (where users performed actions) and negative samples (where users did not perform actions). The intermediary interpolation model bridges the gap between biased recommendation data and unbiased training requirements, allowing the system to use efficiently collected recommendation data while correcting for display bias through the interpolation process.
Solution Approach 2:
The patent transforms the training data by changing the parameter of label generation from direct observation to interpolated prediction. Instead of using raw user action labels directly from recommendation displays, the system generates new prediction labels through the interpolation model that accounts for display probability. This parameter transformation converts biased observed labels into corrected prediction labels that reflect true user preferences independent of display bias.
2Adaptability or versatility
If commodities are displayed based on predicted selection rates, then recommendation personalization is improved, but selection bias increases causing Matthew effect
Solution Approach 1:
The patent inverts the conventional approach by not directly using user action labels from biased recommendation displays as training labels. Instead, it uses the interpolation model to work backwards from observed actions to infer what user preferences would be without display bias. The system inverts the causal relationship: rather than assuming displayed items reflect true preferences, it uses the interpolation model to strip away display bias effects and recover true user preferences from the biased observation data.
3Ease of manufacture
If location bias is present in training data, then data collection simplicity is improved, but model prediction accuracy deteriorates
Solution Approach 1:
The patent extracts and removes the location bias component from the training data through the interpolation model. The interpolation model identifies and separates the bias introduced by display location from the true user preference signal. By taking out the bias component through mathematical interpolation between positive and negative samples, the system retains the simplicity of collecting data from actual recommendation displays while removing the harmful location bias effect that would otherwise degrade model accuracy.
Data Source
AI summary
This application provides a recommendation model training method in the artificial intelligence (AI) field. The training method includes: obtaining a first training sample; processing attribute information of a first user and information about a first recommended object based on an interpolation model, to obtain an interpolation prediction label of the first training sample; and performing training by using the attribute information of the first user and the information about the first recommended object as an input to a recommendation model and using the interpolation prediction label of the first training sample as a target output value of the recommendation model, to obtain a trained recommendation model. According to the technical solutions of this application, impact of training data bias on recommendation model training can be alleviated, and recommendation model accuracy can be improved.