General

Correct the font formatting in all figures

correct figure 7






FROM JACKY
In general, a well formatted study, good work. 

P15: Perhaps including Pred(25) commonly used in software cost estimation? The currently evaluation criteria used are good enough, nevertheless will be convenient for people do not familiar with Win-Tie-Loss calculations. 

Fig 10,11,12,13, if sample size is provided, it will aid the reader, I understand it is already shown in Fig 8. As sample size is important in this study. 

In P19, Fig 12, use Desharnais as an example, some of the MMRE values are too good according to my study. For example in the first box (Uniform h=1), when k=3, MMRE = 0.25 ??? 

If you are using Desharnais dataset with 81 samples, I remember that there are missing values, so they removed them and becomes Deshnarnais 77, and distributed as part of ANGEL's software package (as an example dataset for people to use). 

My understanding is that without classify the dataset into 3 subsets (dev=1, Dev=2, Dev=3) the best you can get is only around 0.65MMRE , no where near 0.25.
Even after classify them into 3 subsets, you will get mostly 0.33, 0.35 MMRE. 

What is not clear to me is the "features" used in the study, as the prediction performance heavily depends on the feature used. 
Understand that perhaps you have included all the features with different weights for different features, but still this is very unlikely to obtain such as good accuracy. 
I am just concerning the results in the paper, correctly me if it indeed produced such results...

sample size
feature
imputation


FROM MENZIES

the discussion around figures1 and 2 should not be so early in the
paper- too much detail early in the paper.
+++ added a new section: Discussion

i would move all of that to a new section after 4 and before 5 called
"discussion" that should be all about "are the above results sane? is
it possible that kernels don't mater?" 
+++DONE

and for that section, the samples should be 10,100,1000
+++DONE: added one sentence why we selected these sample sizes: In this simulation example we chose the instance sizes of $50$ and $100$ on purpose, because, software effort datasets used in our research fall into the range of these numbers.

also, regarding fig7, the dots are still in different places.
reviewers will spasm on this. we spoke about this. please do not give
reviewers any reason to doubt the paper/
+++ gave a new figure and re-wrote the whole thing

lets fix up those 2 things this week. check the result with me, and
lets submit this thing and move on.
---