ClusMatch: Improving Deep Clustering by Unified Positive and Negative Pseudo-Label Learning

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Deep clustering methods show promise but lack annotation, limiting performance.
• A performance gap exists between deep clustering and semi-supervised classification, even with few labels.

Purpose of the Study:

• To bridge the gap between deep clustering and semi-supervised learning.
• To introduce ClusMatch, a unified framework for positive and negative pseudo-label learning in deep clustering.

Main Methods:

• ClusMatch is a pluggable framework adaptable to existing deep clustering techniques.
• It utilizes pre-trained networks for initial predictions and selects high-quality samples for supervised learning.
• A novel unified positive and negative pseudo-label learning strategy is employed for unselected samples, with adaptive thresholding for confidence.

Main Results:

• ClusMatch demonstrated superiority across six widely-used and one large-scale dataset.
• Achieved an average accuracy improvement of 5.4% over the state-of-the-art ProPos method on six datasets.

Conclusions:

• ClusMatch effectively transforms unsupervised clustering into a semi-supervised problem.
• The framework significantly enhances deep clustering performance by incorporating pseudo-labeling strategies.