Now by default, shifted mutual information is non-negative. If the parameter negative_info is set to TRUE, negative shifted mutual information is allowed during the computation when mutual information is inferior to the complexity term. For small dateset with complicated structures, e.g., discrete variables with many levels, allowing for negative shifted mutual information may help identifying weak v-structures related to those discrete variables, as the negative three-point information in those cases will come from the difference between two negative shifted mutual information terms (expected to be negative due to the small sample size). However, under this setting, a v-structure (X -> Z <- Y) in the final graph does not necessarily imply that X is dependent on Y conditioning on Z, As a consequence, the interpretability of the final graph is hindered. In practice, it's advised to keep this parameter as FALSE.
Now by default, shifted mutual information is non-negative. If the parameter
negative_infois set toTRUE, negative shifted mutual information is allowed during the computation when mutual information is inferior to the complexity term. For small dateset with complicated structures, e.g., discrete variables with many levels, allowing for negative shifted mutual information may help identifying weak v-structures related to those discrete variables, as the negative three-point information in those cases will come from the difference between two negative shifted mutual information terms (expected to be negative due to the small sample size). However, under this setting, a v-structure (X -> Z <- Y) in the final graph does not necessarily imply that X is dependent on Y conditioning on Z, As a consequence, the interpretability of the final graph is hindered. In practice, it's advised to keep this parameter asFALSE.