Deep Clustering With a Constraint for Topological Invariance Based on Symmetric InfoNCE.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Mining

Background:

• Deep clustering methods struggle with complex data topologies due to limited prior knowledge.
• Existing state-of-the-art techniques often fail to generalize across both simple and complex data structures.

Purpose of the Study:

• To enhance deep clustering performance on diverse data topologies, including complex ones.
• To address the limitations of current deep clustering algorithms in scenarios with scarce prior information.

Main Methods:

• A novel constraint utilizing symmetric InfoNCE was developed and integrated into deep clustering.
• A new deep clustering method, MIST (Method Integrating Symmetric Topology), was proposed, combining an existing method with the novel constraint.
• Theoretical explanations were provided for the constraint's performance-enhancing capabilities.

Main Results:

• Numerical experiments using MIST demonstrated the effectiveness of the symmetric InfoNCE constraint.
• MIST achieved superior performance compared to other state-of-the-art deep clustering methods across most benchmark datasets.
• The proposed constraint proved beneficial for both noncomplex and complex topology datasets.

Conclusions:

• The symmetric InfoNCE constraint is a valuable addition to deep clustering, improving robustness and performance.
• MIST offers a promising advancement in deep clustering, particularly for datasets with challenging topological features.
• The findings suggest broader applicability of this constraint in unsupervised learning tasks involving complex data structures.