Efficient Image Debiased Contrastive Clustering.

Area of Science:

• Computer Science
• Machine Learning
• Artificial Intelligence

Background:

• E-commerce and social media generate vast image data, challenging real-time clustering and recommendation systems.
• Existing multistage or large-pretrained-model (LPM) clustering methods offer high accuracy but incur significant computational costs and large model sizes.
• Single-stage methods are resource-efficient but struggle with limited feature diversity and accuracy issues due to false positives/negatives.

Purpose of the Study:

• To develop an efficient and lightweight clustering model for image data.
• To address the limitations of existing single-stage and multistage clustering methods.
• To improve accuracy and efficiency in real-time image clustering and recommendation.

Main Methods:

• Proposed Debiased Contrastive Clustering (DCC), an efficient lightweight model.
• Integrated differential augmentations and refined sampling for enhanced feature representation.
• Employed debiased contrastive loss to minimize false negatives and pseudo-labels with consistency regularization to mitigate false positives.

Main Results:

• DCC demonstrated superior performance across seven challenging datasets compared to state-of-the-art (SOTA) methods.
• Achieved higher accuracy, normalized mutual information (NMI), and adjusted Rand index (ARI).
• Exhibited faster convergence and improved overall efficiency without reliance on multistage training or LPMs.

Conclusions:

• DCC effectively balances accuracy and efficiency in image clustering.
• The proposed methods overcome limitations of existing approaches, offering a viable solution for large-scale image data.
• DCC provides a promising direction for real-time clustering and recommendation systems in e-commerce and social media.