Toward Explainable Multiparty Learning: A Contrastive Knowledge Sharing Framework.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Distributed Systems

Background:

• Multiparty learning enables joint model training with decentralized data, but faces challenges like system/statistical heterogeneity and incentive design.
• Existing methods struggle with privacy, data distribution dependence, and communication efficiency.

Purpose of the Study:

• To propose a novel contrastive multiparty learning framework for efficient knowledge refinement and sharing.
• To address limitations of traditional multiparty learning, including privacy concerns and heterogeneity issues.

Main Methods:

• Developed a contrastive multiparty learning framework simulating human cognition and communication for knowledge sharing.
• Introduced an accountable incentive mechanism for enhanced participation.
• Modeled multiparty learning as a many-to-one knowledge-sharing problem, avoiding direct parameter averaging.

Main Results:

• The proposed framework integrates explicit knowledge transparently without privacy disclosure.
• Demonstrated significant model performance improvements across various scenarios and real-world datasets.
• Reduced dependence on data distribution and communication environments.

Conclusions:

• The novel contrastive multiparty learning framework effectively refines and shares knowledge from decentralized data.
• The approach offers a robust solution to heterogeneity and incentive challenges in multiparty learning.
• Achieved superior model performance and efficiency compared to existing methods.