Bayesian Networks in the Absence of Temporal Information - Bayesian Networks in R: With Applications in Systems Biology - page 25

Tabl e 2. 1 Feature matrix for the R packages covered in Sect. 2.3.1

bnlearn catnet deal pcalg gRbase gRain

Discrete data Ye s Ye s Ye s Ye s Ye s Ye s

Continuous data Ye s No Ye s Ye s Ye s No

Mixed data No No Ye s No No No

Constraint-based learning

Ye s No No Ye s No No

Score-based learning Ye s Ye s Ye s No No No

Hybrid learning Ye s No No No No No

Structure manipulation Ye s Ye s No No Ye s No

Parameter estimation

Ye s Ye s Ye s Ye s No No

Prediction Ye s Ye s No No No Ye s

Approximate inference Ye s No No No No Ye s

niques (cross-validation, bootstrap, conditional probability queries, and prediction).
It is also the only package that keeps a clear separation between the structure of a
network and the associated probability distribution, which are implemented as two
different classes of R objects.
deal implements structure and parameter learning using a Bayesian approach
and handles discrete, continuous, and mixed data (assuming a conditional Gaussian
distribution). The network structure is learned with a hill-climbing greedy search
such as the one described in Algorithm 2.2 , with the posterior density of the network
as a score function and random restarts to avoid local maxima.
pcalg provides a free software implementation of the PC algorithm and is specif-
ically designed to estimate and measure causal effects. It handles both discrete and
continuous data and can account for the effects of latent variables on the network.
The latter is achieved through a modified PC algorithm known as Fast Causal Infer-
ence (FCI), first proposed by Spirtes et al. ( 2001 ).
catnet focuses on discrete, static Bayesian networks from a frequentist point
of view. Structure learning is performed in two steps. First, the node ordering of
the graph is learned from the data using simulated annealing; alternatively, a cus-
tom node ordering can be specified by the user. An exhaustive search is performed
among the network structures with the given node ordering, and the exact max-
imum likelihood solution is returned. Parameter learning and prediction are also
implemented. Furthermore, an extension of this approach for mixed data (assuming
a Gaussian mixture distribution) has been recently made available from CRAN in
package mugnet ( Balov , 2011 ).
Packages gRbase ( Højsgaard et al. , 2010 )and gRain ( Højsgaard , 2010 )fallinto
the second category. They focus on manipulating the parameters of the network, on
prediction and on inference, under the assumption that all variables are discrete. Nei-
ther gRbase nor gRain implement any structure or parameter learning algorithm, so
the Bayesian network must be completely specified by the user.