Entire-Space Multi-Task Model for CVR Prediction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing CVR prediction models face challenges such as sparsity of training data and sample selection bias, leading to unsatisfactory training effects due to the sequential dependence of conversion behaviors on clicks and data distribution differences between training and prediction spaces.
Innovation Solution
An entire-space multi-task model with a shared embedding layer and two branches for CTR and CVR predictions is proposed, using a total conversion rate concept to alleviate these issues by modeling both tasks in the entire space and optimizing the training process to avoid gradient conflicts.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional machine learning methods are used to train CVR prediction models, then the model structure is simple, but training data sparsity and sample selection bias lead to unsatisfactory training effects
Solution Approach 1:
The patent merges CTR prediction and CVR prediction into a unified multi-task learning framework. The model shares common layers for feature extraction while having separate output branches for CTR and CVR predictions, allowing the model to leverage patterns from both tasks to improve training effectiveness despite data sparsity.
Solution Approach 2:
The prediction model is designed with multi-functionality to perform both CTR prediction and CVR prediction simultaneously. The shared embedding layers and feature extraction components serve both prediction tasks, making the model more robust and improving training effects by learning universal representations that generalize across different prediction objectives.
2Reliability
If traditional machine learning methods are used to train CVR prediction models, then the model training process is simple, but sample selection bias occurs due to sequential dependence of conversion behaviors on clicks
Solution Approach 1:
The model is segmented into distinct functional components: shared embedding layers for feature representation, separate prediction branches for CTR and CVR, and independent loss function calculations. This segmentation allows the model to handle the sequential dependence issue by treating CTR and CVR as related but distinct prediction tasks, reducing sample selection bias while maintaining manageable training complexity.
Solution Approach 2:
The patent introduces CTR prediction as an intermediary task that bridges the gap between impression data and conversion behavior. By predicting CTR first and using it to inform CVR prediction, the model indirectly addresses the sequential dependence issue, as the CTR branch captures click-related patterns that help reduce bias in the CVR training process.
3Reliability
If separate models are used for CTR and CVR predictions, then each model can be optimized independently, but data sparsity and sample selection bias cannot be effectively addressed
Solution Approach 1:
The patent combines CTR and CVR prediction models into a single unified architecture with shared components. The shared embedding layers and feature extraction modules are jointly trained on both tasks, allowing the model to leverage data from both prediction objectives simultaneously. This merging approach effectively addresses data sparsity and sample selection bias by learning from the combined signal of both tasks, while the modular design keeps the overall structure manageable.
Data Source
AI summary
A target sample including a sample feature, a first label, and a second label is obtained, where a user corresponds to the target sample, the first label indicates whether a target object is clicked on by the user, and the second label indicates whether the user implements a target behavior related to the target object. Model processing is performed on the sample feature using a prediction model with a first branch and first probability and a second branch and second probability. A first loss is determined based on a first label value and the first probability. When a predetermined condition is satisfied, the second loss is determined based on a second label value of the second label and the second probability, and a predicted loss of the target sample is determined based on the first loss and the second loss. Training the prediction model based on the predicted loss.