home - word ranking
Word Ranking
We find a ranking of most discriminative words for all network models using restricted pairwise binary decision trees. For every possible model pair (i,j) for 1<= i<j<=N_mod, where N_mod is the total number of models, we build a binary decision tree, but restricted so that at every level of each tree the same word has to be used for all the trees. At every level the best word is chosen according to the smallest average training loss over all binary trees. The model is not meant to be a substitution to an ordinary multi-class decision tree. It merely represents an algorithm which may be useful to find a fixed number of most discriminative words, for example for visualization of the distributions in a three-dimensional subspace.