evaluate.array {causalDisco}

Evaluate adjacency matrix estimation

Description

Applies several different metrics to evaluate difference between estimated and true adjacency matrices. Intended to be used to evaluate performance of causal discovery algorithms.

Usage

## S3 method for class 'array'
evaluate(est, true, metrics, ...)

Arguments

est

Estimated adjacency matrix/matrices.

true

True adjacency matrix/matrices.

metrics

List of metrics, see details.

...

Further arguments that depend on input type. Currently only list.out is allowed, and only if the first argument is a matrix (see details under Value).

Details

Two options for input are available: Either est and true can be two adjacency matrices, or they can be two arrays of adjacency matrices. The arrays should have shape n * p * p where n is the number of of matrices, and p is the number of nodes/variables.

The metrics should be given as a list with slots $adj, $dir and $other. Metrics under $adj are applied to the adjacency confusion matrix, while metrics under $dir are applied to the conditional orientation confusion matrix (see confusion). Metrics under $other are applied without computing confusion matrices first.

Available metrics to be used with confusion matrices are precision, recall, specificity, FOR, FDR, NPV, F1 and G1. The user can supply custom metrics as well: They need to have the confusion matrix as their first argument and should return a numeric.

Available metrics to be used as "other" is: shd. The user can supply custom metrics as well: They need to have arguments est_amat and true_amat, where the former is the estimated adjacency matrix and the latter is the true adjacency matrix. The metrics should return a numeric.

Value

A data.frame with one column for each computed metric and one row per evaluated matrix pair. Adjacency metrics are prefixed with "adj_", orientation metrics are prefixed with "dir_", other metrics do not get a prefix. If the first argument is a matrix, list.out = TRUE can be used to change the return object to a list instead. This list will contain three lists, where adjacency, orientation and other metrics are reported, respectively.