Deep Adversarial Inconsistent Cognitive Sampling for Multiview Progressive Subspace Clustering.

Area of Science:

• Computer Science
• Machine Learning
• Artificial Intelligence

Background:

• Deep multiview clustering methods show strong performance but neglect sample difficulty labels, leading to suboptimal training.
• Inconsistent difficulty labels across multiview samples pose a significant challenge for existing methods.

Purpose of the Study:

• To propose a novel deep adversarial inconsistent cognitive sampling (DAICS) method for multiview progressive subspace clustering.
• To address the challenge of inconsistent difficulty labels in multiview data for improved clustering accuracy.

Main Methods:

• Developed a deep adversarial inconsistent cognitive sampling (DAICS) approach utilizing a multiview binary classification loss and feature similarity loss.
• Employed an adversarial minimax game to learn a binary classifier and a consistent feature embedding network over difficulty labels.
• Introduced a cognitive sampling strategy with a theoretical guarantee and a golden section mechanism for progressive sample selection based on difficulty.

Main Results:

• The proposed DAICS method effectively handles inconsistent difficulty labels in multiview samples.
• The cognitive sampling strategy and golden section mechanism enable progressive training from easy to difficult samples.
• Experimental results on four real-world datasets show DAICS outperforms state-of-the-art multiview clustering methods.

Conclusions:

• DAICS offers a superior approach to multiview clustering by incorporating difficulty labels and managing inconsistencies.
• The method's ability to progressively train on samples of varying difficulty leads to more efficient and accurate clustering.
• DAICS represents a significant advancement in deep multiview clustering techniques.