B Yang> how did you deal with missing values ? For example, in feature set m1, what did you do for numeric Age value if the member's age is missing ?

Please see the third bullet under "Notation" in A.2 "Features derived from Claim data" for how missing values in Claim data were treated. As for "numeric Age", the documentation refers to [4] for conversion of categorical values to numerical values; however, I just realized that we assigned -1 for missing age, whereas [4] assigned 80.

Thanks for the quick reply. Do your algorithms handle missing values or just use them as is ? For the numeric Age case, do they treat -1 as 'age missing' and do something special, or do they treat it same as other valid age values ?

How about values like avgs and SDs of various fields, for example LengthOfStay. For members who have at least one LengthOfStay value, do you just calculate min/max/avg/whatever while ignoring the missing values ? What about members who have all LengthOfStay values missing ?

Oleg wrote:
"Could you please let us know what do we miss here?"

Good observation about table 1.
Table 1 was generated in order to clarify the paydelay distribution.
The routine used to generate that table contained an error which resulted in incorrect counts.
The correct table is presented below.
Sorry for the confustion, and thanks for finding and reporting this fault.

In the milestone 3 report, the text where month(6-7) is stated should now be changed month(7-8).
We will update this in the milestone 3 report.

Edward.

B Yang> Do your algorithms handle missing values or just use them as is ? For the numeric Age case, do they treat -1 as 'age missing' and do something special, or do they treat it same as other valid age values ? How about values like avgs and SDs of various fields, for example LengthOfStay. For members who have at least one LengthOfStay value, do you just calculate min/max/avg/whatever while ignoring the missing values ? What about members who have all LengthOfStay values missing ?

In our system, the algorithms (RGF, GBDT, random forests, etc.), do not handle missing values. You can't add or compare missing values, so yes, missing values were ignored in (i.e., excluded from) the computation of min/max/avg etc. For the members with all the values missing, min/max/avg etc. were set to zero except for "range" which was set to -1. I hope that this answers all of your questions.

Hi Rie,
I'm trying to use the RGF code you linked to in your paper. When I try to compile the code by running the makefile in os x, I get the error

Dans-Laptop:rgf1.2 dan2$ make
/bin/rm -f bin/rgf
g++ src/tet/driv_rgf.cpp src/com/AzDmat.cpp src/tet/AzFindSplit.cpp src/com/AzIntPool.cpp src/com/AzLoss.cpp src/tet/AzOptOnTree_TreeReg.cpp src/tet/AzOptOnTree.cpp src/com/AzParam.cpp src/tet/AzReg_Tsrbase.cpp src/tet/AzReg_TsrOpt.cpp src/tet/AzReg_TsrSib.cpp src/tet/AzRgf_FindSplit_Dflt.cpp src/tet/AzRgf_FindSplit_TreeReg.cpp src/tet/AzRgf_Optimizer_Dflt.cpp src/tet/AzRgforest.cpp src/tet/AzRgfTree.cpp src/com/AzSmat.cpp src/tet/AzSortedFeat.cpp src/com/AzStrPool.cpp src/com/AzSvDataS.cpp src/com/AzTaskTools.cpp src/tet/AzTETmain.cpp src/tet/AzTETproc.cpp src/com/AzTools.cpp src/tet/AzTree.cpp src/tet/AzTreeEnsemble.cpp src/tet/AzTrTree.cpp src/tet/AzTrTreeFeat.cpp src/com/AzUtil.cpp -Isrc/com -Isrc/tet_tools -O2 -o bin/rgf
src/tet/AzTree.cpp: In member function ‘void AzTree::finfo(AzIFarr*, AzIFarr*) const’:
src/tet/AzTree.cpp:232: error: call of overloaded ‘abs(double&)’ is ambiguous
/usr/include/stdlib.h:146: note: candidates are: int abs(int)
/usr/include/c++/4.2.1/cstdlib:174: note: long long int __gnu_cxx::abs(long long int)
/usr/include/c++/4.2.1/cstdlib:143: note: long int std::abs(long int)
make: *** [all] Error 1

Hopefully it's just a matter of changing the type declarations, but I'm years out of practice in C++.
Thanks for any help,
Dan

Hi DanB,

Interesting. That "abs" is a typo for "fabs", but the two compilers I tested happened to have "double abs(double)", so I missed it. Thanks much for catching this.

Could you please change "abs(w)" in Line 232 of rgf1.2/src/tet/AzTree.cpp to "fabs(w)"?
If you run into further problems, please email to "Contact" in the readme.

Rie

Congratulations to the Milestone 3 winners.
Question for crescendo:

Hi Rie,

In section 4.1 of crescendo_milestone3.pdf (Blending to produce the winning submission), you discuss the approximate ridge regression solution using Leaderboard scores, and note that:

||p(j) - y||^2 can be approximated by n*sj where sj is the Leaderboard score of the j-th run pj.

If the leaderboard scores are root mean squared errors, shouldn't that be n*sj*sj rather than n*sj, or am I confused?

Thanks,

-- Dave Slate

Hi David,

David wrote> shouldn't that be nsjsj rather than n*sj, or am I confused?

Right. That's a typo. That s_j should have a superscript "2".

Rie