Adversarial incomplete multi-view clustering with adaptive contrastive learning.

Area of Science:

• Machine Learning
• Data Science
• Computer Vision

Background:

• Incomplete multi-view data is common due to sensor issues or data loss.
• Existing deep learning methods struggle to balance view contributions and can over-rely on specific features.
• This leads to suboptimal performance in incomplete multi-view clustering tasks.

Purpose of the Study:

• To propose a novel deep incomplete multi-view clustering method named A²CLN.
• To overcome limitations of existing methods in balancing view contributions and feature quality.
• To improve the accuracy and robustness of clustering for incomplete multi-view datasets.

Main Methods:

• Developed an adaptive contrastive learning module to dynamically adjust parameters based on shared information significance.
• Integrated a generative adversarial network (GAN) for adversarial training to enhance latent feature representations.
• A²CLN combines adaptive contrastive learning and adversarial networks for robust incomplete multi-view clustering.

Main Results:

• A²CLN effectively extracts shared information while preserving complementary features from available views.
• Adversarial training via GAN improved the quality of latent feature representations.
• Experiments on six datasets showed A²CLN significantly outperformed existing deep incomplete multi-view clustering methods.

Conclusions:

• A²CLN offers an effective solution for deep incomplete multi-view clustering.
• The adaptive contrastive learning and adversarial network integration enhance clustering structure and feature representation quality.
• The proposed method demonstrates superior performance compared to state-of-the-art techniques.