Duration Prediction Model Training with Dummy Labels for Unclicked Content
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing duration prediction models face challenges with small training data volume, sample selection deviation, and gradient update conflicts due to the imbalance between clicked and unclicked network content, leading to low accuracy in predicting browsing duration.
Innovation Solution
A method involving an auxiliary model to generate dummy duration labels for unclicked samples, followed by training a duration prediction model using these labels along with true duration labels for both clicked and unclicked samples, to improve the model's accuracy in predicting browsing duration.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a duration prediction model is trained using only clicked network content samples, then the model can learn from actual user engagement data, but the training data volume is insufficient and sample selection deviation occurs
Solution Approach 1:
The patent creates dummy duration labels by copying the labeling process from positive samples to negative samples. An auxiliary model generates synthetic duration labels for unclicked samples, allowing the main duration prediction model to learn from both clicked and unclicked content without requiring actual browsing duration data for unclicked items.
Solution Approach 2:
The patent performs preliminary action by pre-processing negative samples through the auxiliary model to generate dummy duration labels before training the main duration prediction model. This preliminary labeling step enables the main model to receive balanced training data in advance, improving training efficiency and accuracy.
2Quantity of substance
If the training data includes both clicked and unclicked network content, then the training data volume increases, but sample selection deviation and gradient update conflict arise due to the imbalance between positive and negative samples
Solution Approach 1:
The patent applies local quality by treating positive and negative samples differently in the training process. Positive samples (clicked content) receive actual duration labels, while negative samples (unclicked content) receive dummy duration labels generated by the auxiliary model. This differentiated approach allows the model to learn appropriate patterns from each sample type without gradient update conflicts.
Solution Approach 2:
The auxiliary model serves as an intermediary between the raw training data and the main duration prediction model. It generates dummy duration labels for negative samples, acting as a mediator that transforms unclicked content into a format suitable for training the main model, thereby reducing sample selection deviation and gradient update conflicts.
3Measurement precision
If an auxiliary model is introduced to generate dummy duration labels for negative samples, then the training data quality improves, but the device complexity increases
Solution Approach 1:
The auxiliary model is designed with multi-functionality, serving multiple purposes: generating dummy duration labels for negative samples, balancing the training data distribution, and reducing sample selection deviation. This universal approach allows a single additional model to address multiple training issues simultaneously, justifying the increased system complexity.
Data Source
AI summary
A training sample subset is acquired from a training sample set, the training sample subset includes a plurality of pieces of sample network content having positive samples corresponding to clicked pieces of sample network content and negative samples corresponding to not clicked pieces of sample network content. A target quantity is determined based on a quantity of the positive samples. By using an auxiliary model that is to be trained, a duration label prediction is performed on selected negative samples of the target quantity obtain dummy duration labels respectively associated with the selected negative samples. True duration labels respectively associated with the plurality of pieces of sample network content are acquired. A duration prediction model and the auxiliary model are trained based on the selected negative samples, the dummy duration labels, the plurality of pieces of sample network content, and the true duration labels.