Choice and redundancy of descriptors

We started with the 48 arithmetical and 170 topological indices from Appendix A. Of these 218 descriptors, 25 ene removed as they are constant within our library 

We searched the remaining 193 indices for pairwise complete correlations. Complete correlation is an equivalence relation. In the following all 19 equivalence classes with more than one element are listed. Most of these complete correlations are due to the special nature of the real libreny. Complete correlations that are generally true are written as 

We used only the first entry from each of these equivalence classes, the other 22 are excluded from further investigation. Thus 171 nonconstant, pairwise incompletely correlated indices remain. The indices twc and are replaced, as before in Section 7.5, 

In this example, various methods of supervised learning are demonstrated. First we considered ABA as a continuous variable, represented by MIC, and obtain predicting functions via regression. As in the sections above, we determined best linear models using OLS and BSS. The following best 5-descriptor model obtained in this manner 

100 random experiments were performed to select best combinations of 5 out of 171 pseudodescriptors to fit 51 pseudoobservations. The result was mhrR = 0.56013, stdev = 0.04772. Note the high mhrR value resulting from a 5-descriptor combination to be selected from a large descriptor pool. 

---