Server Device for Federated Learning Node Value Integration
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Horizontal federated learning for Gradient Boosting Decision Trees (GBDT) faces inefficiencies due to the high computation and communication requirements of encrypting and transmitting gradient information, making it difficult to perform effectively.
Innovation Solution
A server device and client devices collaborate by acquiring and integrating node value information from each client, determining node values without gradient information transmission, and learning the decision tree based on shared data structures.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If gradient information is encrypted and transmitted for horizontal federated learning, then model training accuracy is improved, but computation quantity increases enormously
Solution Approach 1:
The patent extracts and transmits only the essential node value information from the decision tree structure, rather than transmitting complete gradient information. This selective extraction maintains model training accuracy while significantly reducing computation quantity by eliminating unnecessary cryptographic operations on full gradient data.
Solution Approach 2:
The patent uses node value information as a simplified copy or representation of the complete gradient information. This copying approach allows the server to obtain sufficient training signals without requiring the computationally expensive encryption and transmission of full gradient vectors, thus reducing computation quantity while preserving accuracy.
2Measurement precision
If gradient information is encrypted and transmitted for horizontal federated learning, then model training accuracy is improved, but communication amount increases
Solution Approach 1:
The patent extracts only the critical node value parameters from the decision tree structure for transmission. This extraction reduces communication amount by sending only essential information needed for model training, while the server can reconstruct sufficient training signals without receiving complete encrypted gradient information.
Solution Approach 2:
The patent transmits node value information as a compact copy that represents the essential training signals. This copying mechanism reduces communication overhead compared to transmitting full encrypted gradient information, while maintaining model training accuracy through the use of these condensed representations.
3Measurement precision
If gradient information is transmitted multiple times for determining node values, then decision tree learning accuracy is improved, but communication frequency increases
Solution Approach 1:
The patent extracts node value information that encapsulates multiple gradient signals into a single transmission. This extraction reduces communication frequency by consolidating multiple rounds of gradient exchanges into fewer transmissions, while maintaining decision tree learning accuracy through the aggregated node value information.
4Measurement precision
If complete gradient information is transmitted for each client, then model training completeness is improved, but data privacy protection becomes weaker
Solution Approach 1:
The patent extracts only node value information from the complete gradient information, which preserves model training completeness while reducing privacy risk. The extracted node values contain sufficient training signals without exposing the full gradient information that could reveal sensitive client data patterns.
Solution Approach 2:
The patent uses node value information as a privacy-preserving copy that represents gradient information without containing the full detail. This copying approach maintains model training completeness while weakening the privacy risk associated with transmitting complete gradient information across the network.
Data Source
AI summary
A server device includes an acquisition unit that acquires, from each of a plurality of client devices, information representing a value of a node constituting a decision tree, the information being determined based on the learning data held by the own device of each client device, and a determination unit that determines the value of the node constituting the decision tree by integrating the acquired results. The decision tree is learned by determination of the value of each node constituting the decision tree by the determination unit.