Soft Neighbors Supported Contrastive Clustering.

Area of Science:

• Machine Learning
• Data Mining
• Computer Vision

Background:

• Deep clustering methods often use contrastive or non-contrastive learning.
• Contrastive methods typically compare positive and negative sample pairs.
• Existing approaches overlook inter-sample relationships, causing class collisions and degraded performance.

Purpose of the Study:

• To propose a novel soft neighbor-supported contrastive clustering method.
• To address the limitations of hard sample treatment in existing contrastive clustering.

Main Methods:

• Introduced a 'perception radius' to quantify sample-neighbor similarity.
• Designed a two-level soft neighbor loss for local and global neighborhood relationships.
• Incorporated a cluster-level loss for compact and separated distributions.
• Implemented a pseudo-label refinement strategy to mitigate false negatives.

Main Results:

• The proposed method significantly improves deep clustering performance.
• Demonstrated superiority over existing methods on benchmark datasets.
• The soft neighbor approach effectively captures inter-sample relationships.

Conclusions:

• Soft neighbor-supported contrastive clustering offers a superior approach to deep clustering.
• The method effectively handles inter-sample relationships, leading to enhanced clustering accuracy.
• This work provides a new direction for developing robust deep clustering algorithms.