This virtual base class represents a score for causal inference; it is used
in the causal inference functions ges, gies and
simy.
The fields of Score are mainly of interest for users who aim at
deriving an own class from this virtual base class, i.e., implementing an own
score function.
.nodes:Node labels. They are passed to causal inference methods by default to label the nodes of the resulting graph.
decomp:Indicates whether the represented score is decomposable (cf. details). At the moment, only decomposable scores are supported by the implementation of the causal inference algorithms; support for non-decomposable scores is planned.
pp.dat:List representing the preprocessed input data; this is typically a statistic which is sufficient for the calculation of the score.
.pardag.class:Name of the class of the parametric DAG model
corresponding to the score. This must name a class derived from
ParDAG.
c.fcn:Only used internally; must remain empty for (user
specified) classes derived from Score.
new("Score",
data = matrix(1, 1, 1),
targets = list(integer(0)),
target.index = rep(as.integer(1), nrow(data)),
nodes = colnames(data),
...)
dataData matrix with \(n\) rows and \(p\) columns. Each row corresponds to one realization, either interventional or observational.
targetsList of mutually exclusive intervention targets that have been used for data generation.
target.indexVector of length \(n\); the \(i\)-th entry
specifies the index of the intervention
target in targets under which the \(i\)-th row of data
was measured.
nodesNode labels
...Additional parameters used by derived (and non-virtual) classes.
Note that since Score is a virtual class, its methods cannot be called
directly, but only on derived classes.
local.score(vertex, parents, ...)For decomposable scores, this function calculates the local score of a vertex and its parents. Must throw an error in derived classes that do not represent a decomposable score.
global.score.int(edges, ...)Calculates the global score of a DAG, represented as a list of in-edges: for each vertex in the DAG, this list contains a vector of parents.
global.score(dag, ...)Calculates the global score of a DAG,
represented as object of a class derived from ParDAG.
local.fit(vertex, parents, ...)Calculates a local model fit
of a vertex and its parents, e.g. by MLE.
The result is a vector of parameters whose
meaning depends on the model class; it matches the convention used in the
corresponding causal model (cf. .pardag.class).
global.fit(dag, ...)Calculates the global MLE of a DAG,
represented by an object of the class specified by .pardag.class.
The result is a list of vectors, one per vertex, each in the same format
as the result vector of local.mle.
Alain Hauser (alain.hauser@bfh.ch)
Score-based structure learning algorithms for causal inference such as
Greedy Equivalence Search
(GES, implemented in the function ges), Greedy Interventional
Equivalence Search (GIES, implemented in the function gies) and
the dynamic programming approach of Silander and Myllymäki (2006)
(implemented in the function simy) try to find the DAG model which
maximizes a scoring criterion for a given data set. A widely-used scoring
criterion is the Bayesian Information Criterion (BIC).
The virtual class Score is the base class for providing a scoring
criterion to the mentioned causal inference algorithms. It does not
implement a concrete scoring criterion, but it defines the functions that
must be provided by its descendants (cf. methods).
Knowledge of this class is only required if you aim to implement an own
scoring criterion. At the moment, it is recommended to use the predefined
scoring criteria for multivariate Gaussian data derived from Score,
GaussL0penIntScore and
GaussL0penObsScore.
T. Silander and P. Myllymäki (2006). A simple approach for finding the globally optimal Bayesian network structure. Proceedings of the 22nd Conference on Uncertainty in Artificial Intelligence (UAI 2006), 445--452
ges, gies, simy,
GaussL0penIntScore,
GaussL0penObsScore