I really need help to understand how the function interact.gbm (gbm package) can compute Friedman's H statistic to assess the relative strengh of interaction effects.
In this Friedman's paper http://arxiv.org/pdf/0811.1679.pdf (section 8.1, page 20), we can see how Friedman computes this statistic.
The question is why does he consider partial dependance function as additive ? It's totally not the case for me : the range of a partial dependance function F_j is the same as the range of the predictive function F.
Thank you !
