NNI algorithms return wrong number of rows not imputed

[Jari Häkkinen]

The estimate function of the NNI algos are supposed to return the number of not imputed rows. The current implementation says that a row was not imputed as soon as one value on that row was not imputed. Is this what we want? I think the that no imputed rows should be rows where no value was imputed ... this also implies that rows with all weights equal to zero will become no impute rows.

We are keeping the current implementation. The expected return value is zero, the return value can be used to detect poor imputation. The estimate() function doc will be improved. Also a test for the return values should be added to the test suite.

[Peter]

I think as it is what we wanted. I guess the idea was that if the return is zero, everything is fine, i.e., all values have been imputed. If it returns anything else, you have a problem. I don't remember why we wanted that information though; was it a simple error check?

[Jari Häkkinen]

The doc on he estimate() function is

/// Function doing the imputation.
///
/// @return number of rows not imputed

but it should say something smarter like expect a zero return value for a good imputation ... or should we remove the return value and forget about it?

[Peter]

Would it speed up the function if we made it void?

I think we can keep it, and modify the docs as you suggested.

[Jari Häkkinen]

(In [1726]) Fixes #476. Improved estimate() doc and added test on estimate() return value.