Closed j-chou closed 9 years ago
OK-but I'm still hoping you'll try something simple that always makes a zero score the best: Unless all three are zero, then toss out that quartet. I think A*Exp[-bx], where A is the biggest you want a score to b, and b is a parameter that makes a bigger separation of scores near zero, is simple and worth trying. I'm attaching a pdf to remind you of the behavior of this function for some choices of this function for various A and b. Please add this to your list for the experiment.
On Fri, Apr 10, 2015 at 5:36 PM, j-chou notifications@github.com wrote:
Sorry, I posted on the wrong github page.
Shashank and I were talking about the svd scores to weights conversion, and shashank suggested the following scheme: Let R=1/e^(svd_max-svd_med). Then weight(qmin)=e^(R (1-svd_min/svd_overall_max)) weight(qmed)=e^(R (1-svd_med/svd_overall_max)) weight(q_max)=e^(R*(1-svd_max/svd_overall_max)) where svd_overall_max is the global maximum svd score. This will fix problems with svd_min = 0 and prevent the weights from being too large or too small. Let's call this new scheme H. In the python file we will need a list of all svd scores and compute the max of this list before assigning any weights to the quartets.
— Reply to this email directly or view it on GitHub https://github.com/agupta0905/AGBsvdquartets/issues/4.
Ruth Davidson Department of Mathematics, University of Illinois Urbana-Champaign http://www.math.uiuc.edu/~redavid2/
Of course x would be the svd score, so best score for example for transformation 30*Exp(-100x) would be 30, and if svd_min = 0 get q_max = 30, vs. svd_max (worst) = 0.002 means q_min (worst) = 24.56
Depend on what values you are used to seeing for svd_min, svd_max you could make best choice.
On Fri, Apr 10, 2015 at 6:08 PM, Ruth Davidson redavids@ncsu.edu wrote:
OK-but I'm still hoping you'll try something simple that always makes a zero score the best: Unless all three are zero, then toss out that quartet. I think A*Exp[-bx], where A is the biggest you want a score to b, and b is a parameter that makes a bigger separation of scores near zero, is simple and worth trying. I'm attaching a pdf to remind you of the behavior of this function for some choices of this function for various A and b. Please add this to your list for the experiment.
On Fri, Apr 10, 2015 at 5:36 PM, j-chou notifications@github.com wrote:
Sorry, I posted on the wrong github page.
Shashank and I were talking about the svd scores to weights conversion, and shashank suggested the following scheme: Let R=1/e^(svd_max-svd_med). Then weight(qmin)=e^(R (1-svd_min/svd_overall_max)) weight(qmed)=e^(R (1-svd_med/svd_overall_max)) weight(q_max)=e^(R*(1-svd_max/svd_overall_max)) where svd_overall_max is the global maximum svd score. This will fix problems with svd_min = 0 and prevent the weights from being too large or too small. Let's call this new scheme H. In the python file we will need a list of all svd scores and compute the max of this list before assigning any weights to the quartets.
— Reply to this email directly or view it on GitHub https://github.com/agupta0905/AGBsvdquartets/issues/4.
Ruth Davidson Department of Mathematics, University of Illinois Urbana-Champaign http://www.math.uiuc.edu/~redavid2/
Ruth Davidson Department of Mathematics, University of Illinois Urbana-Champaign http://www.math.uiuc.edu/~redavid2/
best choice for A and b I mean. If biggest score you EVER see is 0.0002 maybe you want b = 150 or something.
Shashank and I decided on the phone that if the biggest SVD score you have ever seen is 0.01, please set A = 10 and b = 100, then the best score will be 10 and the worst score will be 3.7.
On Fri, Apr 10, 2015 at 6:13 PM, Ruth Davidson redavids@ncsu.edu wrote:
Of course x would be the svd score, so best score for example for transformation 30*Exp(-100x) would be 30, and if svd_min = 0 get q_max = 30, vs. svd_max (worst) = 0.002 means q_min (worst) = 24.56
Depend on what values you are used to seeing for svd_min, svd_max you could make best choice.
On Fri, Apr 10, 2015 at 6:08 PM, Ruth Davidson redavids@ncsu.edu wrote:
OK-but I'm still hoping you'll try something simple that always makes a zero score the best: Unless all three are zero, then toss out that quartet. I think A*Exp[-bx], where A is the biggest you want a score to b, and b is a parameter that makes a bigger separation of scores near zero, is simple and worth trying. I'm attaching a pdf to remind you of the behavior of this function for some choices of this function for various A and b. Please add this to your list for the experiment.
On Fri, Apr 10, 2015 at 5:36 PM, j-chou notifications@github.com wrote:
Sorry, I posted on the wrong github page.
Shashank and I were talking about the svd scores to weights conversion, and shashank suggested the following scheme: Let R=1/e^(svd_max-svd_med). Then weight(qmin)=e^(R (1-svd_min/svd_overall_max)) weight(qmed)=e^(R (1-svd_med/svd_overall_max)) weight(q_max)=e^(R*(1-svd_max/svd_overall_max)) where svd_overall_max is the global maximum svd score. This will fix problems with svd_min = 0 and prevent the weights from being too large or too small. Let's call this new scheme H. In the python file we will need a list of all svd scores and compute the max of this list before assigning any weights to the quartets.
— Reply to this email directly or view it on GitHub https://github.com/agupta0905/AGBsvdquartets/issues/4.
Ruth Davidson Department of Mathematics, University of Illinois Urbana-Champaign http://www.math.uiuc.edu/~redavid2/
Ruth Davidson Department of Mathematics, University of Illinois Urbana-Champaign http://www.math.uiuc.edu/~redavid2/
Ruth Davidson Department of Mathematics, University of Illinois Urbana-Champaign http://www.math.uiuc.edu/~redavid2/
Sorry, I posted on the wrong github page.
Shashank and I were talking about the svd scores to weights conversion, and shashank suggested the following scheme: Let R=1/e^(svd_max-svd_med). Then weight(qmin)=e^(R(1-svd_min/svd_overall_max)) weight(qmed)=e^(R(1-svd_med/svd_overall_max)) weight(q_max)=e^(R*(1-svd_max/svd_overall_max)) where svd_overall_max is the global maximum svd score. This will fix problems with svd_min = 0 and prevent the weights from being too large or too small. Let's call this new scheme H. In the python file we will need a list of all svd scores and compute the max of this list before assigning any weights to the quartets.