We found that keeping all numeric variables as numeric and avoiding any quantile binning had a very direct and negative effect on the model performance. The overall score was far lower in the numeric case than in the quantile binning case: AUC: 0.81 for all numeric versus AUC: 0.88 for QB.
Looking at the convergence graph for the All Numeric model, it appears that the algorithm converged much more slowly than it had for the quantile binning model. It obviously had not converged after 50 passes, so we increased the number of passes to 100. We also noticed that in the All Numeric case, the best learning rate was equal to 0.01, whereas in the quantile binning model, the best learning ...
