How does eigen_centrality() handle missing weights when using the weight option? Does it treat NA as 0? Thanks.
This question turned out to be more difficult than I imagined.
In principle the vector gets passed on to the C core or igraph. It is not immediately clear to me what NA from R is translated to in C, apparently R uses a separate representation of NA (i.e. different from NaN). When reading through the code though it seem as if the check on NA is not done correctly:
if (!is.null(weights) && any(!is.na(weights))) {
weights <- as.numeric(weights)
} else {
weights <- NULL
}
What seems to be required instead is all(!is.na(weights)) or !any(is.na(weights)). I believe this is an error. Could you confirm this @Gabor? Or am I overlooking something here?
Does the R interface pass on the binary representations of NA to the C core as-is? @vtraag you are saying that NAs are detected and the weight vector changed appropriately, right? (Not just passed on.)
On the types of CPUs we use, the “NaN” value has more than one possible representation. My understanding is that R uses a specific NaN representation for NA. If this is so, and if the binary representation were passed to C as-is, then as far as the C core is concerned, all NAs would really just be NaNs. IMO that would make no sense—it should not be allowed.
What you say makes sense to me but I am not experienced with R at all.
AFAIK, the weight vector was intended to be set to NULL if any of the values are NA. This should be corrected in the R interface (it can be easily corrected in the types-R.def file I believe).
I’m not exactly sure what the representation of NA is, but indeed R seems to use a specific NaN value for representing the NA value of R. In the translation from R vectors to igraph_vector_t (done by R_SEXP_to_vector) no translation of any values occur at all. In other words, as it is now the specific value of NA gets passed to the igraph C core, which indeed does not make any sense. We do not perform any checks on this in the C core either by the way.
I agree that results with weights containing any NA are totally nonsense if NAs are just NaNs (there is a table on here summarizing the corresponding c++ type for NA in Rcpp. Rcpp does interpret NAs as NaNs for doubles but for the raw c interface the conversion does not seem to be documented).
If that’s the case, I will open an issue on rigraph repo.
Sorry for resurrecting an old thread, but I think this definitely wasn’t the intention. It is way too easy to accidentally introduce a NA value in a numeric vector in R; also, it is pretty commonplace to use NA as a placeholder for unknown values. Numeric weight vectors (potentially containing thousands of values) should not be silently ignored if there is a NA value somewhere in the middle; I think that the correct behaviour in this case is to convert NA to NaN when transferring them over to the C layer, and make the C layer handle them accordingly (or blow up with an error if the C layer cannot handle them).