HetVis: A Visual Analysis Approach for Identifying Data Heterogeneity in Horizontal Federated Learning.

Area of Science:

• Machine Learning
• Data Visualization
• Privacy-Preserving Technologies

Background:

• Horizontal federated learning (HFL) allows distributed clients to collaboratively train models while preserving data privacy.
• Data heterogeneity across clients is a significant challenge impacting the quality of HFL models.
• Investigating data heterogeneity is difficult due to security concerns and model complexity.

Purpose of the Study:

• To develop a visual analytics tool, HetVis, for clients to explore and understand data heterogeneity in HFL.
• To provide methods for identifying and summarizing data heterogeneity issues.
• To aid client analysts in recognizing and addressing heterogeneity.

Main Methods:

• Developed HetVis, a visual analytics tool, based on requirement analysis.
• Identified data heterogeneity by comparing global HFL model predictions with local stand-alone model predictions.
• Employed context-aware clustering for inconsistent records to summarize heterogeneity.
• Designed novel visualizations and comparison techniques for HFL heterogeneity analysis.

Main Results:

• HetVis enables clients to explore data heterogeneity by comparing model prediction behaviors.
• Context-aware clustering provides a summarized view of data heterogeneity.
• Novel visualizations effectively highlight heterogeneity issues in HFL.
• Case studies demonstrated HetVis's utility in understanding various heterogeneity types.

Conclusions:

• HetVis is an effective tool for clients to investigate data heterogeneity in HFL.
• The developed visualization techniques aid in identifying and understanding heterogeneity.
• Expert reviews and comparative studies validate the effectiveness of HetVis in improving HFL model quality.