Differentiable self-supervised clustering with intrinsic interpretability.

Area of Science:

• Machine Learning
• Data Mining
• Artificial Intelligence

Background:

• Self-supervised clustering methods discover latent structures without labels but often lack interpretability.
• Existing approaches struggle to provide transparent insights into the data clustering process.

Purpose of the Study:

• To propose a differentiable self-supervised clustering method with intrinsic interpretability (DSC2I).
• To develop an interpretable data clustering mechanism using differentiable programming.

Main Methods:

• Designed a differentiable mutual information measurement for training neural networks with analytical gradients.
• Developed an interpretable clustering mechanism by converting clustering objectives into neural networks.
• Integrated representation learning and interpretable clustering within a unified optimization framework.

Main Results:

• The proposed DSC2I method achieves effective and interpretable data clustering.
• Demonstrated superior performance compared to 16 existing clustering approaches in extensive experiments.
• Learned discriminative and compact representations through analytical gradient-based training.

Conclusions:

• DSC2I offers a transparent and interpretable approach to self-supervised clustering.
• The method successfully combines representation learning and clustering in a unified, self-supervised manner.
• DSC2I advances the field by providing inherent interpretability in unsupervised learning.