Classification From Pairwise Similarities/Dissimilarities and Unlabeled Data via Empirical Risk Minimization

Area of Science:

• Machine Learning
• Computer Science
• Data Science

Background:

• Pairwise similarities and dissimilarities are often more accessible than full data labels in real-world classification.
• Existing empirical risk minimization methods can use pairwise similarities but not dissimilarities.
• Semisupervised clustering methods can use both but often require strong, potentially limiting, geometrical assumptions.

Purpose of the Study:

• To derive an unbiased estimator for classification risk using both similarities and dissimilarities with unlabeled data.
• To address the limitations of existing methods in handling pairwise dissimilarities.
• To provide a more comprehensive approach for semi-supervised learning utilizing all available pairwise information.

Main Methods:

• Developed an empirical risk minimization approach incorporating both pairwise similarities and dissimilarities.
• Derived an unbiased estimator for classification risk.
• Established a theoretical estimation error bound for the proposed method.

Main Results:

• The proposed method successfully utilizes both similarities and dissimilarities with unlabeled data for risk estimation.
• Theoretical analysis confirmed an estimation error bound.
• Experimental results demonstrated the practical utility and effectiveness of the new method.

Conclusions:

• The novel empirical risk minimization method offers a robust way to leverage all pairwise information (similarities and dissimilarities) in classification.
• This approach overcomes limitations of prior methods and semisupervised clustering techniques.
• The method shows practical applicability and theoretical soundness for semi-supervised learning tasks.