Core Concepts
Developing a privacy-preserving framework for horizontal federated XGBoost without sharing gradients, improving communication efficiency.
Abstract
Introduction: Discusses the need for training XGBoost in federated learning due to privacy concerns.
Existing Works: Focus on neural networks in FL, limited exploration for other ML models like XGBoost.
Challenges: Horizontal vs. vertical settings in federated XGBoost, difficulties in optimal split conditions.
Proposed Solution: Introduces FedXGBllr framework with learnable learning rates to address privacy and communication efficiency.
Methodology: Formulates intuitions, facilitates them through a one-layer 1D CNN, and develops the FedXGBllr framework.
Experiments: Extensive evaluations show comparable performance to state-of-the-art methods and reduced communication overhead by factors ranging from 25x to 700x.
Results: Outperforms or matches accuracy of SimFL and centralized baselines on classification datasets; achieves comparable MSE on regression datasets.
Ablation Studies: Demonstrates interpretability of the one-layer 1D CNN model coupled with high performance.
Communication Overhead Comparison: Significantly lower communication overhead compared to SimFL, saving costs by factors ranging from 25x to 700x.
Stats
"Our approach achieves performance comparable to the state-of-the-art method."
"Effectively improves communication efficiency by lowering both communication rounds and overhead by factors ranging from 25x to 700x."