Calculation of CLR transformation

[Vlasovets]

the currently used function looks as follows:

def transform_features(
    features: pd.DataFrame, transformation: Str = "clr", coef: float = 0.5
) -> pd.DataFrame:
    if transformation == "clr":
        X = features.values
        #null_set = X <= 0.0 # just ignore zero replacement for the sake of experiment
        #X[null_set] = coef
        X = np.log(X)
		____________
		#bug is here
        X = (X.T - np.mean(X, axis=1)).T
		# one could change to 
		X = (X - np.mean(X, axis=0))
		____________

        return pd.DataFrame(
            data=X, index=list(features.index), columns=list(features.columns)
        )

    else:
        raise ValueError(
            "Unknown transformation name, use clr and not %r" % transformation
        )

a small test example, if check the first two values by hand, you will see the mistake

df = pd.DataFrame(np.random.randint(0,10,size=(3, 3)), columns=list('ABC'))
transform_features(df)

Cheers,
Oleg

[Leo-Simpson]

Hello Oleg,
Thank you for the pull request, and for "taking the lead" on the project somehow.

So it might be yes, then there might be something I do not understand. Shouldn't the mean be performed on each row of matrix X (if row i contains the features of sample i ) ?

[Vlasovets]

oh, if row i contains the features of sample i", so your X is an (nxp) matrix where n - samples, p - features then axis=1 is correct since we need a geometric mean per sample
thanks for the clarification and sorry if I confused you.