Spectral Embedded Deep Clustering.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• Unsupervised clustering of unlabeled data is a fundamental challenge in machine learning.
• Existing deep neural network clustering methods often rely on restrictive assumptions, such as uniform cluster balance.
• The need for flexible and robust clustering algorithms applicable across diverse data domains is critical.

Purpose of the Study:

• To propose a novel deep neural network-based clustering method for unlabeled datasets.
• To develop a method that directly groups data in its original feature space without prior assumptions on cluster distribution.
• To enhance clustering robustness and applicability across various data types.

Main Methods:

• Utilizes a deep neural network to define a conditional discrete probability distribution as a statistical model.
• Employs a strategy of initially estimating cluster labels for high-density data points.
• Applies semi-supervised learning using estimated labels and remaining unlabeled data for model training.

Main Results:

• The proposed method successfully groups unlabeled data into a specified number of clusters.
• Demonstrates effectiveness without requiring assumptions of uniform cluster balance, unlike prior methods.
• Exhibits robustness against outliers and applicability to data represented as feature vectors.
• Numerical experiments on five datasets confirm the method's superior performance compared to existing approaches.

Conclusions:

• The developed deep neural network clustering method offers a flexible and robust alternative to existing techniques.
• Its ability to handle non-uniform cluster distributions and outliers makes it suitable for a wider range of applications.
• The method shows significant promise for advancing unsupervised learning and data analysis.