PCBN (version 0.1.1)

fit_copulas: Fit the copulas of a PCBN given data

Description

Fit the copulas of a PCBN given data

Fit all possible orders given a DAG

Usage

fit_copulas(
  data,
  DAG,
  order_hash,
  familyset = c(1, 3, 4, 5, 6),
  familyMatrix = NULL,
  e,
  verbose = 1,
  method = "mle"
)
fit_all_orders(
  data,
  DAG,
  familyset = c(1, 3, 4, 5, 6),
  e,
  score_metric = "BIC",
  verbose = 1
)

Value

fit_copulas returns the fitted PCBN, with an additional element called metrics which is a named vector of length 3 with names c("logLik", "BIC", "AIC"), where \(AIC = - 2 * logLik + 2 * nparameters\)

and \(BIC = - 2 * logLik + log(n) * nparameters\), for a sample size n and nparameters is the number of parameters.

fit_all_orders returns a list containing:

• best_fit the PCBN which is the best according to the metric score_metric.

• fitted_list the list of all fitted PCBNs.

• metrics the matrix of metrics (logLik, BIC, AIC). Each row i of this matrix corresponds to a PCBN with a different set of parents' orderings, and corresponds to element i of fitted_list.

Arguments

data

data frame

DAG

Directed Acyclic Graph

order_hash

hashmap of parental orders

familyset

vector of copula families

familyMatrix

matrix of families, with named rows and columns. This should be used if the copula families are known/fixed. This overrides familyset.

e

environment containing all the hashmaps

verbose

if 0, don't print anything. If verbose >= 1, print information about the simulation procedure.

method

indicates the estimation method ("mle" for maximum likelihood estimation or "itau" of inversion of Kendall's tau).

score_metric

name of the metric used to choose the best order. Possible choices are logLik, AIC and BIC.

See Also

BiCopCondFit which this function wraps.

Examples

DAG = create_empty_DAG(3)
DAG = bnlearn::set.arc(DAG, 'U1', 'U3')
DAG = bnlearn::set.arc(DAG, 'U2', 'U3')

order_hash = r2r::hashmap()
order_hash[['U3']] = c("U1", "U2")

fam = matrix(c(0, 0, 1,
               0, 0, 1,
               0, 0, 0), byrow = TRUE, ncol = 3)

tau = 0.2 * fam

my_PCBN = new_PCBN(
  DAG, order_hash,
  copula_mat = list(tau = tau, fam = fam))

mydata = PCBN_sim(my_PCBN, N = 5)
e = default_envir()

result = fit_copulas(data = mydata, DAG = DAG,
                     order_hash = order_hash,
                     familyset = 1, e = e)

result_all_orders = fit_all_orders(data = mydata, DAG = DAG,
                                   familyset = 1, e = e)

# The two fitted PCBNs are:
print(result_all_orders$fitted_list[[1]])
print(result_all_orders$fitted_list[[2]])
# and the metrics matrix is:
print(result_all_orders$metrics)

# The PCBN corresponding to the true order U1 < U2 is usually better
# than the second one. This Will be more clear with a bigger sample size.