Transductive methods for the distributed ensemble classification problem.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Science

Background:

• Distributed ensemble classification presents challenges due to the absence of common labeled training data.
• Existing methods often rely on fixed combination rules, which may not be optimal.
• Proprietary or legacy classifiers and diverse sensing modalities complicate ensemble design.

Purpose of the Study:

• To develop new methods for distributed ensemble classification without shared labeled data.
• To optimize the classifier aggregation function using unlabeled test data.
• To improve classification accuracy in scenarios with prior mismatch.

Main Methods:

• A transductive approach is employed, optimizing the combining rule on unlabeled test data.
• Maximum Likelihood (ML) objectives are proposed, using expectation-maximization for iterative adjustment.
• An information-theoretic method is introduced, designed to outperform ML methods and handle redundancies.

Main Results:

• ML methods yield probabilistic aggregation forms, adjusted for class prior mismatches.
• The information-theoretic method shows superior performance, handles classifier redundancies, and addresses missing classes.
• All proposed methods demonstrated improved classification accuracy on benchmark data compared to fixed rules under prior mismatch conditions.

Conclusions:

• The developed transductive methods offer significant improvements for distributed ensemble classification.
• The information-theoretic approach provides a robust and versatile alternative to ML-based methods.
• These findings are particularly relevant for real-world applications with heterogeneous data sources and no shared labeled data.