Recommendation Model Training with Segmented Historical Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing recommendation systems suffer from low accuracy due to sample selection bias during model training, as they primarily use historical display data without incorporating diverse historical recommendation data, leading to incomplete training samples.
Innovation Solution
A data processing method that extracts object attribute information and historical behavior data from log data, acquires historical recommendation data, and trains a recommendation model using a combination of historical display data, click data, and non-click data to enhance model training accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If only historical display data is used for model training, then the training process is simple, but the model accuracy is low due to sample selection bias
Solution Approach 1:
The patent segments historical recommendation data into three distinct subsets: historical display data, historical click data, and historical non-click data. This segmentation allows the model to learn from different types of user interactions separately, reducing sample selection bias while maintaining manageable training complexity through structured data organization.
Solution Approach 2:
The patent merges multiple types of historical recommendation data (display data, click data, non-click data) into a unified training dataset. This combination enriches the training samples with diverse user behavior patterns, improving model accuracy by capturing both positive (click) and negative (non-click) feedback signals alongside display information.
2Measurement precision
If diverse historical recommendation data is incorporated, then model training accuracy improves, but data processing complexity increases
Solution Approach 1:
The patent divides the diverse historical recommendation data into distinct segments (display data, click data, non-click data) with specific processing procedures for each. This segmentation strategy enables targeted data preprocessing and feature extraction for each data type, improving training accuracy while controlling processing complexity through modular data handling.
Solution Approach 2:
The patent performs preliminary processing of diverse historical recommendation data before model training, including data cleaning, feature extraction, and labeling. This preliminary action prepares the data in advance, reducing the computational burden during actual training and making the complex data processing more manageable through systematic preprocessing steps.
Data Source
AI summary
Provided is a data processing method and apparatus and a storage medium. Object attribute information, historical behavior data and historical display data corresponding to the object attribute information are extracted from historical log data. Historical recommendation data corresponding to the object attribute information is acquired from a historical recommendation information base. The historical recommendation data is searched for first historical recommendation data which is the same as the historical display data. Second historical recommendation data is obtained according to the historical display data, historical behavior data and the first historical recommendation data. A preset recommendation model is trained by using the second historical recommendation data and the third historical recommendation data to obtain a trained preset recommendation model. Upon reception of first identity attribute information, recommendation data corresponding to the first identity attribute information is determined based on the trained preset recommendation model.