select.parameter: Tuning parameter selection

In sparsebnUtils: Utilities for Learning Sparse Bayesian Networks

Description

Choose the best DAG model according to the criterion described in Fu and Zhou (2013) (Section 3.4).

Usage

select.parameter(x, data, type = "profile", alpha = 0.1)

Arguments

x	sparsebnPath object.
data	sparsebnData containing the original data.
type	either "profile" or "full", default is profile.
alpha	tuning parameter for selection between 0.05 and 0.1, default is 0.5 (see equation (11) in Fu and Zhou (2013)).

Details

A sparsebnPath objects represents a solution path which depends on the regularization parameter lambda. Model selection is usually based on an estimated prediction error, and commonly used model selection methods include the Bayesian information criterion (BIC) and cross-validation (CV) among others. It is well-known that these criteria tend to produce overly complex models in practice, so instead we employ an empirical model selection criterion that works well in practice. As lambda is decreased and thus the model complexity increases, the log-likelihood of the estimated graph will increase. An increase in model complexity, which is represented by an increase in the total number of predicted edges, is desirable only if there is a substantial increase in the log-likelihood. In order to select an optimal parameter, this method computes successive difference ratios between the increase in log-likelihood and the increase in number of edges and balances these quantities appropriately. For specific details, please see Section 3.4 in Fu and Zhou (2013).

sparsebnUtils documentation built on Jan. 27, 2021, 9:05 a.m.