User Behavior Prediction Training With Delay-Weighted Labels
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing user behavior prediction models suffer from inaccurate labeling due to delayed user interactions, leading to incorrect classification of positive samples as negative samples, which affects prediction accuracy.
Innovation Solution
A method for training a user behavior prediction model that involves setting two labels for each sample, determining a delay status based on these labels, and performing weighted combination of prediction losses to adjust model parameters, thereby accounting for different importance of samples with varying delay statuses.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If the latest data generated within a short time are used to update the machine learning model, then the model can be updated frequently, but the prediction accuracy deteriorates due to delayed user behavior feedback
Solution Approach 1:
The patent introduces a time delay parameter τ to distinguish between immediate user behaviors and delayed behaviors. By parameterizing the time delay and creating different label types (first label for current behavior, second label for delayed behavior within duration τ), the model can differentiate between samples with different delay statuses, thereby maintaining high prediction accuracy while using recent data for frequent updates
Solution Approach 2:
The patent segments the label information into two distinct labels: the first label indicating current user behavior and the second label indicating delayed user behavior within a specific time duration. This segmentation allows the training process to treat samples with different delay statuses differently through weighted combination, resolving the contradiction between using latest data and maintaining accuracy
2Device complexity
If all samples are treated equally during training, then the training process is simple, but the prediction accuracy deteriorates due to incorrect labeling of delayed positive samples as negative samples
Solution Approach 1:
The patent applies local quality by assigning different weights to different samples based on their specific characteristics (delay status). Samples are categorized into different types (current negative samples, delayed positive samples, etc.) and each type receives appropriate weighting. This localized differentiation ensures that the training process accounts for the specific quality of each sample, improving prediction accuracy while maintaining reasonable training complexity
3Productivity
If the time interval between user behavior and target object exposure is not considered, then the data processing is efficient, but the label accuracy deteriorates due to positive samples being incorrectly marked as negative samples
Solution Approach 1:
The patent performs preliminary action by pre-defining a time duration τ and pre-categorizing samples based on their delay status before training. The second label is assigned in advance to indicate whether the user behavior occurred within the time duration τ from target object exposure. This preliminary classification ensures that label accuracy is maintained while allowing efficient processing during the actual training phase
Data Source
Figure 1
Figure 2~3
Figure 4
AI summary
The disclosure provides methods and apparatuses for training a user behavior prediction model. A sample set generated through streaming is obtained, where any sample includes a sample feature, a first label, and a second label. A sample feature of each sample is input into the model, to obtain a prediction result about whether a corresponding user performs specific behavior. Each corresponding prediction loss is determined based on the prediction result and a first label value. A sample category that indicates a delay status and to which each sample belongs is determined based on the first label value and a second label value, and a weight value of each sample is determined based on the sample category. Weighted combination is performed on each prediction loss corresponding to each sample based on the weight value, and a parameter of the model is adjusted based on an obtained combined loss.