Toward Projected Clustering With Aggregated Mapping.

Area of Science:

• Machine Learning
• Data Science
• Computer Vision

Background:

• Projected clustering is fundamental to deep clustering.
• Existing models may suffer from representation degeneration and over-fitting.
• Novel frameworks are needed to enhance deep clustering capabilities.

Purpose of the Study:

• To propose a novel projected clustering framework.
• To address the over-fitting issue in representation learning for deep clustering.
• To improve the performance of deep clustering models.

Main Methods:

• Developed an aggregated mapping technique combining projection learning and neighbor estimation.
• Introduced a self-evolution mechanism to aggregate sub-clusters and mitigate over-fitting.
• Theoretically analyzed the degeneration risk in representation learning.
• Demonstrated unsupervised projection function selection with linear and non-linear examples.

Main Results:

• The proposed framework obtains a clustering-friendly representation.
• The self-evolution mechanism effectively alleviates over-fitting risks.
• Ablation experiments confirm the theoretical analysis and the effectiveness of neighbor aggregation.
• Significant improvements in deep clustering performance were observed.

Conclusions:

• The novel projected clustering framework enhances deep clustering by learning robust representations.
• The self-evolution mechanism is crucial for preventing degeneration and improving model generalization.
• The proposed method offers a promising direction for advancing deep clustering techniques.