Clustering Analysis via Deep Generative Models With Mixture Models.

Area of Science:

• Machine Learning
• Data Mining
• Pattern Recognition

Background:

• Traditional clustering algorithms struggle with complex feature interdependencies in latent spaces.
• Deep generative models like VAEs and GANs excel at learning latent representations for unsupervised tasks.

Purpose of the Study:

• To propose a novel deep generative clustering approach.
• To enhance clustering and generation performance, especially with outlier data.
• To validate the model's effectiveness through clustering analysis and sample generation.

Main Methods:

• Developed a novel clustering approach integrating Wasserstein GAN with gradient penalty (WGAN-GP) and variational autoencoder (VAE) with a Gaussian mixture prior.
• Formulated the WGAN-GP generator using samples from the VAE's probabilistic decoder.
• Introduced a variant using a Student's-t mixture prior for improved robustness against outliers.

Main Results:

• The proposed deep generative model demonstrated stable training and improved clustering accuracy compared to state-of-the-art methods.
• Experiments validated the model's effectiveness in both clustering analysis and realistic sample generation.
• The Student's-t mixture prior variant showed robust performance in the presence of data outliers.

Conclusions:

• The novel WGAN-GP and VAE-based deep generative model offers a superior approach to clustering.
• The model achieves stable training, enhanced accuracy, and realistic sample generation.
• The Student's-t mixture prior variant provides a robust solution for clustering with outlier data.