Federated Learning Model Training Using Adaptive Splitting
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Federated learning models struggle to fully utilize data due to heterogeneous or homogeneous data distribution, leading to poor model performance and data loss, as existing technologies often discard data and require labels to be present in one party, which is impractical.
Innovation Solution
The method involves mixing horizontal and vertical federated learning modes, using vertical boosting trees for heterogeneous data and horizontal boosting trees for homogeneous data, allowing for lossless data utilization and improving model performance by automatically selecting the appropriate learning mode based on data distribution.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If existing federated learning technologies are used, then data privacy is protected, but data utilization is insufficient leading to poor model performance
Solution Approach 1:
The patent segments the federated learning process into two distinct modes: horizontal federated learning for homogeneous data and vertical federated learning for heterogeneous data. This segmentation allows the system to selectively apply the appropriate learning mode based on data distribution characteristics, thereby improving model performance while maintaining privacy protection. The boosting tree structure is also segmented into multiple weak classifiers that are trained in parallel across different parties.
Solution Approach 2:
The patent introduces dynamic adaptability by automatically selecting between horizontal and vertical federated learning modes based on the data distribution characteristics. The system dynamically adjusts the learning approach according to whether the data is homogeneous or heterogeneous, eliminating the need for manual configuration and optimizing model performance for different data scenarios while preserving privacy constraints.
2Device complexity
If data is discarded to simplify the learning process, then computational complexity is reduced, but data loss increases leading to poor model performance
Solution Approach 1:
The patent applies local quality by treating different types of data differently based on their distribution characteristics. For homogeneous data, horizontal federated learning is applied; for heterogeneous data, vertical federated learning is applied. This localized approach ensures that each data type is processed with the most appropriate method, maximizing data utilization without unnecessarily increasing overall system complexity.
Solution Approach 2:
The patent creates a universal federated learning framework that can handle both homogeneous and heterogeneous data through a single system. The framework incorporates both horizontal and vertical learning modes, allowing it to universally process different data types without requiring separate systems or discarding data. This multi-functionality eliminates data loss while managing complexity through unified architecture.
3Stability of the object's composition
If labels are required to be present in one party, then label consistency is ensured, but data utilization is limited and model performance deteriorates
Solution Approach 1:
The patent introduces an intermediary mechanism where the server coordinates between multiple parties, each holding different data and labels. The server aggregates information from all parties without requiring any single party to possess all labels, thereby maintaining label consistency through coordinated computation while fully utilizing data from all participants. This intermediary role enables collaborative learning across data silos.
Data Source
AI summary
In a method for training a federated learning model, a server obtains a target split mode corresponding to a training node in response to determining that the training node satisfies a preset splitting condition. The server notifies a client to perform, based on the target split mode, node splitting. The server performs a next round of training by taking a left subtree node generated by performing the node splitting as a new training node until an updated training node does not satisfy the preset splitting condition. The server performs a next round of training by taking another non-leaf node of the boosting tree as a new training node. The server stops training and generates a target federated learning model in response to determining that a node dataset of the plurality of boosting trees is empty.