Deep Autoencoding Topic Model With Scalable Hybrid Bayesian Inference.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Natural Language Processing

Background:

• Traditional topic models often lack flexibility and interpretability.
• Scalable inference for complex generative models remains a challenge.

Purpose of the Study:

• To develop a flexible and interpretable deep autoencoding topic model (DATM).
• To enable scalable posterior inference and efficient latent representation learning for document analysis.
• To enhance discriminative power for jointly modeling documents and labels.

Main Methods:

• Developed a Deep Autoencoding Topic Model (DATM) with a hierarchy of gamma distributions.
• Implemented topic-layer-adaptive stochastic gradient Riemannian MCMC for scalable posterior inference.
• Proposed a Weibull upward-downward variational encoder for efficient local latent representation inference.
• Introduced supervised DATM for joint document and label modeling.

Main Results:

• Demonstrated the efficacy of DATM on unsupervised document analysis tasks.
• Showcased the scalability of DATM on big corpora.
• Validated the effectiveness of supervised DATM in enhancing discriminative power for classification tasks.

Conclusions:

• DATM provides a flexible and interpretable framework for advanced document analysis.
• The proposed inference methods ensure scalability for large datasets.
• Supervised DATM effectively improves latent representation discriminability for related tasks.