Federated Learning Model Integration via Data Similarity Weighting
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Federated learning systems struggle to effectively integrate models considering inference target data, particularly in healthcare settings where data distribution differences across client terminals are not adequately addressed, leading to suboptimal performance of integrated models.
Innovation Solution
A federated learning system where client terminals learn models using local data sets and a server integrates these models by calculating similarities between data samples and learning data sets, adjusting for client IDs and objective variables, and generating integrated analysis models through repeated learning and integration processes.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If federated learning integrates models from multiple client terminals without considering data distribution differences, then model integration can be achieved, but inference performance deteriorates due to heterogeneity in learning data across clients
Solution Approach 1:
The patent applies local quality by calculating similarity between each client's learning data and the input data to determine data-specific weights. Each client contributes differently to the integrated model based on how well their local data distribution matches the inference target, thereby resolving the contradiction between efficient model integration and accurate inference by adapting the integration strategy to local data characteristics.
Solution Approach 2:
The patent changes the parameter of model integration by introducing similarity-based weighting instead of uniform averaging. The integration process dynamically adjusts weights based on data distribution similarity metrics, transforming the integration mechanism from a static equal-weight approach to a dynamic adaptive approach that preserves inference accuracy while maintaining integration efficiency.
2Device complexity
If similarity calculation is performed only by the server, then calculation load on client terminals is reduced, but communication overhead increases due to transmitting all learning data to the server
Solution Approach 1:
The patent extracts the critical similarity calculation function from the server and implements it locally on each client terminal. By taking out this computational task from the centralized server, the system reduces communication overhead and data transmission requirements while maintaining the ability to perform accurate similarity assessments for model integration.
Solution Approach 2:
The patent applies preliminary action by having client terminals calculate similarities between their own learning data and input data before model integration. This preliminary local calculation avoids the need to transmit all learning data to the server, reducing communication overhead while preparing the necessary information for weighted model integration in advance.
3Ease of manufacture
If learning models are trained independently on each client terminal without weighting, then training simplicity is maintained, but model integration quality deteriorates due to unaddressed data distribution differences
Solution Approach 1:
The patent implements feedback by using similarity calculation results to dynamically adjust model integration weights. The system measures the similarity between client data distributions and the inference target, then feeds this information back into the integration process to weight models appropriately, thereby maintaining training simplicity while significantly improving integrated model performance and reliability.
Solution Approach 2:
The patent introduces dynamics into the model integration process by making weights adaptive rather than static. The integration mechanism dynamically adjusts the contribution of each client's model based on real-time similarity assessments, allowing the system to maintain simple independent training while achieving reliable integrated performance through dynamic weight adjustment.
Data Source
AI summary
A server inputs an input data set into a similarity calculation model, calculates a first similarity between the input data set and learning data sets, each client terminal inputs a data sample to the similarity calculation model to calculate a second similarity between the data sample and the learning data sets, executes a similarity calculation between the first and second similarity and outputs a learning weight, learns a first analysis model that calculates a predicted value of the first objective variable from the first explanatory variable for the input data set based on the first analysis model, the first explanatory variable and the first objective variable of the learning data set, and the learning weight, and transmits a first learning result. The server generates a second analysis model that integrates the first learning result and calculates a predicted value of the first objective variable from the first explanatory variable.