Multi-Task Learning for Compositional Data via Sparse Network Lasso.

Area of Science:

• Statistics
• Machine Learning
• Bioinformatics

Background:

• Multi-task learning (MTL) improves generalization by sharing information across tasks.
• Compositional data, with components summing to one, present unique challenges for standard MTL due to interdependencies.
• Existing MTL methods are not directly applicable to compositional data.

Purpose of the Study:

• To develop a new multi-task learning method specifically designed for compositional data.
• To address the limitations of existing MTL approaches when applied to data with proportional components.
• To enable the extraction of latent structures and key variables from compositional datasets.

Main Methods:

• Proposed a multi-task learning approach using sparse network lasso regularization.
• Utilized a symmetric log-contrast model for regression with compositional covariates.
• Treated each sample as a distinct task within the multi-task learning framework.

Main Results:

• The proposed method successfully extracts latent clusters and identifies relevant variables in compositional data.
• Demonstrated improved prediction accuracy compared to existing methods in simulation studies.
• Validated effectiveness through application to real-world gut microbiome data.

Conclusions:

• The novel sparse network lasso-based MTL method is effective for compositional data analysis.
• Leveraging sample relationships significantly enhances prediction accuracy for compositional datasets.
• The method offers a powerful tool for exploring complex biological data, such as microbiome compositions.