Boosted network classifiers for local feature selection.

Area of Science:

• Network analysis
• Computational biology
• Machine learning

Background:

• Network feature selection models commonly assume sparsity, where only a small network subset influences a phenomenon.
• This sparsity assumption is often enforced using regularized models like LASSO.
• However, many real-world networks contain highly correlated modules, making the sparsity assumption potentially inappropriate.

Purpose of the Study:

• To introduce two novel optimization strategies: boosted expectation propagation (BEP) and boosted message passing (BMP).
• To evaluate the performance of BEP and BMP against traditional network-regularized logistic regression models.
• To demonstrate the impact of the sparsity assumption on network classifier accuracy in the presence of correlated network structures.

Main Methods:

• Developed BEP and BMP as ensemble methods that combine models based on local network features.
• These methods directly utilize network structure to estimate parameters for a network classifier.
• Unlike sparse models, BEP and BMP do not assume sparsity but seek a weighted average of network features, emphasizing useful ones for classification.

Main Results:

• BEP and BMP were compared with network-regularized logistic regression on simulated and real biological networks.
• Results indicated that assuming sparsity adversely affects accuracy and feature selection power in networks with highly correlated structures.
• The proposed BEP and BMP methods showed superior performance in such scenarios.

Conclusions:

• The sparsity assumption can be detrimental to network classifier performance when dealing with highly correlated network modules.
• Boosted expectation propagation (BEP) and boosted message passing (BMP) offer effective alternatives that leverage network structure without assuming sparsity.
• These novel methods enhance accuracy and feature selection capabilities, particularly in complex biological networks.