Deep divergence-based approach to clustering.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Mining

Background:

• Deep learning for clustering is an emerging field, with challenges in designing effective loss functions for unsupervised representation learning.
• Traditional clustering methods often struggle with high-dimensional data and scalability.

Purpose of the Study:

• To develop a novel deep clustering network capable of learning representations and cluster structures simultaneously.
• To introduce a new loss function incorporating information-theoretic divergence and geometric regularization for improved clustering performance.

Main Methods:

• Proposed a deep clustering network architecture.
• Developed a novel loss function combining information-theoretic divergence measures with geometric regularization constraints.
• Evaluated the network on synthetic and real-world datasets.

Main Results:

• The proposed network achieved competitive performance compared to state-of-the-art deep clustering methods.
• The method demonstrated good scalability for large datasets.
• The approach effectively avoids degenerate clustering structures through geometric regularization.

Conclusions:

• The novel deep clustering network offers a promising approach for unsupervised learning tasks.
• The proposed information-theoretic loss function with geometric regularization enhances clustering quality and stability.
• This method provides an effective, pre-training-free solution for deep clustering.