Federated Learning Domain Aggregation for Fairness and Personalization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Federated learning systems face issues of unfairness and lack of personalization due to some client computing devices not contributing to model training, leading to incomplete feature spaces and models that are not optimized for individual client needs.
Innovation Solution
Implementing domain-specific aggregation of local updates by categorizing client devices into subsets based on characteristics, using dimensionality reduction and clustering, and employing queue management and transfer learning to ensure all updates are considered and models are personalized.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If client computing devices are not categorized into domains, then the system is simpler to implement, but the model lacks personalization and fairness for different client types
Solution Approach 1:
The patent segments client computing devices into multiple domains (e.g., mobile devices, wearable devices, IoT devices) based on their characteristics. This segmentation enables domain-specific aggregation of local updates, allowing the system to personalize models for different client types while maintaining manageable complexity through structured categorization
2Reliability
If all client updates are aggregated equally, then the global model is simpler to compute, but updates from inactive or non-contributing clients skew the model training
Solution Approach 1:
The patent applies local quality by aggregating local updates differently for different domains. Each domain has its own aggregation mechanism that weights updates based on domain-specific criteria, ensuring that inactive or non-contributing clients in certain domains do not unfairly skew the global model, while still incorporating valuable updates from active clients
3Adaptability or versatility
If domain-specific aggregation is implemented, then model personalization and fairness improve, but the computation time and resources increase
Solution Approach 1:
By segmenting clients into domains, the system can parallelize the aggregation process across different domains, reducing overall computation time. Each domain can be processed independently, and the results can be combined efficiently, making the domain-specific aggregation scalable and time-efficient
Data Source
AI summary
A processor may receive local updates for a current global model from a plurality of client computing devices, each local update being determined by one of the plurality of client computing devices based on a local dataset on the client computing device. The processor may determine a domain-specific aggregate of the local updates from each subset of the plurality of client computing devices. The processor may determine an updated global model based on the domain-specific aggregate of the local updates for each subset of the plurality of client computing devices.