Remark on Linear Scaling

Normalization reduces the information linearly against the individual descriptor instead of against the complete data set. Several authors have mentioned this as a problem, in particular, concerning the use of descriptors with neural networks. This can be overcome by using linear scaling methods, which use factors that are usually related to the complete set of data. 

However, scaling has two general drawbacks. Each data set needs its own set of scaling factors that is, the data set must be investigated in advance a fact that is 

---