Domain Adaptation via Pseudo-Label Bridge for Healthcare ML
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Enhancing the generalization ability of machine learning models when adapting from a source domain to a target domain, especially in healthcare scenarios where access to the source domain data is restricted and data distributions differ.
Innovation Solution
A method involving obtaining a source model trained to generate labels for data from the source domain, generating predicted labels for target domain data using the source model, and training a target model based on these labels and true target domain labels to enhance knowledge transfer and adaptability.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a machine learning model is trained on source domain data to achieve high accuracy, then the model performance on source domain data is improved, but the model's ability to generalize to target domain data deteriorates when data distributions differ
Solution Approach 1:
The patent introduces an intermediary domain (pseudo-label domain) that acts as a bridge between the source domain and target domain. By generating pseudo-labeled target domain data and training an intermediary model on this synthetic data, the system creates a transitional learning path that helps the model adapt to target domain distribution without direct access to真实 target labels, thereby resolving the contradiction between source domain accuracy and target domain generalization
Solution Approach 2:
The patent performs preliminary actions by first training a source model on source domain data, then using this source model to generate pseudo-labeled target domain data before training the final target model. This preliminary training and pseudo-label generation process prepares the model in advance for target domain adaptation, improving generalization ability while maintaining the structured learning approach
2Reliability
If access to source domain data is restricted to protect data privacy and security, then data security is improved, but the ability to adapt models to new domains deteriorates
Solution Approach 1:
The patent creates copies of source domain knowledge by generating pseudo-labeled target domain data that captures the essential patterns and relationships from the source domain. Instead of directly accessing or transferring source domain data, the system creates synthetic copies through the source model's predictions on target domain data, enabling domain adaptation while maintaining data security and privacy
Solution Approach 2:
The source model acts as an intermediary that transfers knowledge from the source domain without requiring direct access to source domain data. The model learns patterns from source domain training and then applies this knowledge to generate pseudo-labeled target domain data, serving as a mediator that enables adaptation while preserving data security boundaries
3Measurement precision
If a model is trained exclusively on target domain data with true labels, then the training accuracy on target domain data is improved, but the process becomes inefficient when target domain labeled data is limited
Solution Approach 1:
The patent merges source domain pseudo-labeled data with target domain truly labeled data to create a combined training dataset. This merging approach allows the model to leverage both the quantity of source domain knowledge (through pseudo-labels) and the quality of target domain labels, improving training efficiency and accuracy when target domain labeled data is limited by combining advantages from both domains
Data Source
AI summary
Embodiments of the present disclosure relate to a method, system, and computer program product for domain adaptation. According to the method, a source model of a source domain is obtained, where the source model is trained to generate a label indicating a predicted category of data from the source domain. A training sample from a target domain is obtained, where the training sample comprises training data from the target domain and a true label indicating a true category of the training data from the target domain. A first label is generated for the training data by using the source model. The first label indicates a predicted category of the training data. A target model of the target domain is trained based on the training data, the true label and the first label.