Mixture-of-Experts Variational Autoencoder for clustering and generating from similarity-based representations on...

Area of Science:

• Computational biology
• Machine learning
• Data science

Background:

• Clustering high-dimensional data is challenging.
• Deep Clustering methods offer flexibility for complex datasets.

Purpose of the Study:

• Introduce a novel generative clustering model, MoE-Sim-VAE.
• Enable learning of multi-modal data distributions and generation of realistic data.
• Improve clustering performance on complex, high-dimensional datasets.

Main Methods:

• Developed the Mixture-of-Experts Similarity Variational Autoencoder (MoE-Sim-VAE).
• Utilized a Variational Autoencoder (VAE) with a Mixture-of-Experts (MoE) decoder architecture.
• Encouraged a Gaussian mixture distribution in the latent space to capture data similarities.

Main Results:

• MoE-Sim-VAE effectively learns multi-modal data distributions.
• The model demonstrates high efficacy and efficiency in generating realistic data.
• Achieved superior clustering performance on MNIST, single-cell RNA-sequencing, and mass cytometry datasets compared to baselines and competitors.

Conclusions:

• MoE-Sim-VAE is a powerful generative clustering model for high-dimensional data.
• The model's architecture facilitates automatic learning of data modes and accurate similarity representation.
• Demonstrated state-of-the-art performance across various benchmark and real-world biological datasets.