Deep Clustering Analysis via Dual Variational Autoencoder With Spherical Latent Embeddings.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• Deep generative models are increasingly used for unsupervised learning tasks, particularly in extracting meaningful latent embeddings from data.
• Existing variational autoencoder (VAE)-based deep clustering methods often utilize Gaussian mixture models (GMMs) as priors, which may limit latent space representation.

Purpose of the Study:

• To propose a novel deep clustering method that leverages a variational autoencoder (VAE) with spherical latent embeddings.
• To enhance the capabilities of unsupervised clustering through improved latent space modeling and data representation.

Main Methods:

• A novel VAE-based clustering method employing a von Mises-Fisher mixture model prior instead of a GMM prior, resulting in spherical latent embeddings.
• Utilization of a dual VAE structure to enforce reconstruction constraints on latent embeddings and their noise counterparts, embedding data into a hyperspherical latent space.
• Introduction of an augmented loss function for enhanced model robustness through self-supervised learning via mutual guidance between original and augmented data.

Main Results:

• The proposed method achieves effective deep generative clustering by learning spherical latent embeddings.
• The use of the von Mises-Fisher mixture model prior allows for principled control over decoder capacity and latent embedding utilization.
• The dual VAE structure and augmented loss function contribute to a robust and self-supervised clustering process.

Conclusions:

• The novel VAE-based deep clustering method with spherical latent embeddings demonstrates superior performance compared to state-of-the-art methods on benchmark datasets.
• The proposed approach offers a principled and robust framework for unsupervised clustering using deep generative models.
• The method's effectiveness is validated, and source code is publicly available for further research and application.