How can I adjust classifier to the scale of the other I have 2 classifiers with different scales. How can I adjust one classifier to the scale of the other without loss of quality? 
On the scatter plot we have 2 solutions plotted (x1, x2) against each other. I believe using linear combination as a link function is not appropriate in this particular case.
Help is much appreciated.
 A: John Tukey described a simple method that works well in practice.  Pick three representative points within the scatterplot: one near each end and one in the middle.  Experiment with Box-Cox (power + log) transforms of either or both variables, applying them only to the coordinates of the three representatives.  (Tukey did this by hand; a spreadsheet speeds up the calculation.)  You want to make the slope between the left point and the middle point closely equal the slope between the middle and the right point.  Applying the resulting transformation(s) will linearize the plot.  (It basically has to: if you can get these points to line up, everything will line up.)
We can hope for homoscedasticity, too.  Your plot seems to exhibit a little more scatter on the right hand side.  This suggests choosing Box-Cox parameters (powers) less than 1.
For example, I can eyeball three points on your plot at (0.2, 0.05), (0.5, 0.2), and (0.7, 0.4).  Using just powers (which are Box-Cox transformations up to an affine change) I get slopes of 0.70 and 0.71 when leaving the x-coordinate unchanged but taking the 0.3 power of the y-coordinate.  The 0.3 power isn't much fun--one would prefer the 1/2 or maybe the 1/3 power--so I also tried taking the square root (1/2 power) of x and the logarithm (0 "power") of y.  The slopes are now 5.33 and 5.35: exactly equal within the imprecision of my coordinate estimates.
This suggests you should consider using $\left(\sqrt{x_1}, \log{x_2} \right)$ for your variables.  If this makes the scatter around the resulting linear scatterplot a little uneven (heteroscedastic), go back and try less severe powers, such as 1 and 1/3.  When you're done you can apply affine transformations (rescaling and recentering) to get the re-expressed variables in the range $[0, 1/2]$ if you really need that.
Reference:
Tukey, J, EDA.  Addison-Wesley (1977).
