ICAFS: Inter-Client-Aware Feature Selection for Vertical Federated Learning.

Area of Science:

• Machine Learning
• Data Science
• Artificial Intelligence

Background:

• Vertical Federated Learning (VFL) enables collaborative model training on data partitioned across clients.
• Feature Selection (FS) is critical in VFL due to distributed, distinct feature subsets across clients.
• Existing FS methods in VFL often neglect crucial inter-client feature interactions, limiting model performance.

Purpose of the Study:

• To develop an effective feature selection (FS) method for Vertical Federated Learning (VFL) that accounts for inter-client feature interactions.
• To introduce a novel multi-stage ensemble approach named ICAFS for enhanced FS in VFL.
• To improve prediction accuracy in VFL by addressing limitations of intra-client focused FS.

Main Methods:

• Introduced ICAFS, a multi-stage ensemble approach for feature selection in VFL.
• Employed conditional feature synthesis and multiple learnable feature selectors.
• Facilitated ensemble FS using synthetic embeddings, bypassing private gradient sharing.

Main Results:

• ICAFS demonstrated superior performance compared to state-of-the-art methods in prediction accuracy.
• Experiments on multiple real-world datasets validated the effectiveness of the proposed method.
• The approach enables model training with refined embeddings derived from real data.

Conclusions:

• ICAFS offers an effective solution for feature selection in Vertical Federated Learning by considering inter-client feature interactions.
• The proposed method enhances model performance without compromising data privacy through private gradient sharing.
• ICAFS represents a significant advancement in VFL, improving prediction accuracy through sophisticated feature selection.