Entropic Graph-based Posterior Regularization.

Area of Science:

• Machine Learning
• Computational Biology

Background:

• Graph smoothness objectives are successful in semi-supervised learning but underutilized in unsupervised generative models.
• Probabilistic models often lack mechanisms to enforce similarity between posterior distributions of related variables.

Purpose of the Study:

• To introduce a new class of entropic graph-based posterior regularizers for unsupervised generative models.
• To develop an efficient inference and parameter learning algorithm for these regularizers.
• To apply the method to computational biology, specifically genomic data analysis.

Main Methods:

• Defined entropic graph-based posterior regularizers to encourage similar posterior distributions for nearby variables.
• Developed a three-way alternating optimization algorithm with closed-form updates for inference and parameter learning.
• Algorithm updates are linear in graph degree, exhibit monotone convergence, and are parallelizable.

Main Results:

• The proposed method outperforms existing graph-based regularization techniques on a synthetic problem.
• It also surpasses comparable strategies for long-range interactions using existing approximate inference methods.
• Significant improvements were observed in predicting genomic activity when integrating 3D genome interaction data.

Conclusions:

• The novel regularizers effectively augment probabilistic models for unsupervised learning.
• The efficient optimization algorithm facilitates practical application.
• The method demonstrates substantial utility in computational biology for genomic data annotation and prediction.