Hello Everyone,
I have been researching on the quality metrics been used by SAP PA. and Predictive Power (Ki) model performance metric.' Ki 'which is quality metrics is actually Somers'd . But why it has been used ?
Predictive Power (Ki) corresponds to the proportion of information contained in the target variable that the explanatory variables are able to explain.
Somers' delta is a nonparametric measure of the strength and direction of association that exists between an ordinal dependent variable and an ordinal independent variable.
It is not confirmed that all the variable will ordinal ??
Can anyone have an explanation why this quality metric is been used ?
Hi Yashi,
1. SAP BusinessObjects Predictive Analytics uses KI as a performance indicator. KI is 2*AUC - 1 and AUC is a widely used non-parametric order statistic measure. The choice of [-1, +1] values for KI instead of [0 ; 1] for AUC is that it is always easier for an non-advanced user to understand that -1 is bad and +1 is good 😉. Otherwise, optimizing a model based on KI or AUC is the same.
2. KI or AUC , roughly speaking, measures how higher score values map to positive target values (the opposite is also true, lower scores should map to negative cases). The input variables are not always ordinal, but when their categories are ordered by their target profit (percentage of positive cases inside each category), their encoding is.
3. I looked at the wikipedia definition of Somers-D - it seems to be equal to 2*AUC - 1 for logistic regression , which corresponds to the KI formula. Not sure though that this is always the case for any distribution.
Hi Yashi, can you please elaborate why you are making a relation between the KI and Somers' delta? Thanks & regards Antoine
Add a comment