Cross-View Representation Learning-Based Deep Multiview Clustering With Adaptive Graph Constraint.

Area of Science:

• Machine Learning
• Data Mining
• Computer Vision

Background:

• Deep multiview clustering analyzes data with multiple modalities and features.
• Autoencoder (AE)-based methods excel at feature extraction but face challenges with cross-view information and complex local structures.
• Existing methods often use Kullback-Leibler (KL) divergence, leading to less discriminative clusters.

Purpose of the Study:

• To propose two novel AE-based deep multiview clustering algorithms: AG-DMC and ADG-DMC.
• To address limitations in cross-view information integration, local structure preservation, and cluster discrimination.
• To enhance the performance of multiview clustering for complex datasets.

Main Methods:

• AG-DMC employs a cross-view representation learning model using cascaded latent representations and an entropy-regularized adaptive graph constraint.
• ADG-DMC further incorporates an adversarial learning mechanism as the clustering loss for improved discrimination.
• Both methods utilize autoencoders for inherent feature extraction.

Main Results:

• The proposed AG-DMC and ADG-DMC algorithms demonstrated superior performance on eight real-world datasets.
• AG-DMC effectively learns view-consensus features and preserves local data structures.
• ADG-DMC significantly enhances cluster discrimination through adversarial learning.

Conclusions:

• The developed algorithms offer significant improvements over existing advanced multiview clustering techniques.
• The novel approaches effectively handle cross-view information and complex local structures.
• The findings suggest a promising direction for deep multiview clustering research and applications.