Cost-based feature selection for network model choice.

Area of Science:

• Computational Biology
• Machine Learning
• Statistical Modeling

Background:

• Feature selection is crucial for machine learning and Bayesian computation, especially with large, noisy datasets.
• Computational cost of feature calculation is a significant, often overlooked, factor, particularly in network analysis.

Purpose of the Study:

• To develop and evaluate cost-aware feature selection methods for network model selection.
• To reduce the computational burden of identifying informative features in network models.

Main Methods:

• Adapted nine existing feature selection methods to incorporate feature computation costs.
• Utilized pilot simulations on smaller networks to inform feature selection for larger network models.
• Applied the developed approaches to yeast protein interaction networks.

Main Results:

• Reduced computational cost by two orders of magnitude for network models without significant loss in classification accuracy.
• Achieved a 50-fold reduction in computational cost using pilot simulations, maintaining classification accuracy.
• Successfully identified the best-fitting duplication divergence model for three yeast protein interaction networks.

Conclusions:

• Cost-aware feature selection is an effective strategy for network model selection, balancing computational efficiency and accuracy.
• The proposed methods offer practical solutions for reducing computational costs in complex network analyses.
• The findings provide insights into the evolutionary dynamics of yeast protein interaction networks.