[scikit-learn] Confidence and Prediction Intervals of Support Vector Regression

Sebastian Raschka mail at sebastianraschka.com
Wed Mar 1 21:44:13 EST 2017


Hi, Raga,

> 1. Just to make sure I understand correctly, using the .632+ bootstrap method, the ACC_lower and ACC_upper are the lower and higher percentile of the ACC_h,i distribution?

phew, I am actually not sure anymore … I think it’s the percentile of the ACC_boot distribution, similar to the “classic” bootstrap but where ACC_boot got computed from weighted ACC_h,i and ACC_r,i

>  2. For regression algorithms, is there a recommended equation for the no-information rate gamma?


Sorry, can’t be of much help here; I am not sure what the equivalent of the no-information rate for regression would be ...



> On Mar 1, 2017, at 5:39 PM, Raga Markely <raga.markely at gmail.com> wrote:
> 
> Thanks a lot, Sebastian! Very nicely written.
> 
> I have a few follow-up questions:
> 1. Just to make sure I understand correctly, using the .632+ bootstrap method, the ACC_lower and ACC_upper are the lower and higher percentile of the ACC_h,i distribution?
> 2. For regression algorithms, is there a recommended equation for the no-information rate gamma?
> 3. I need to plot the confidence interval and prediction interval for my Support Vector Regression prediction (just to clarify these intervals, please see an analogy from linear model on slide 14: http://www2.stat.duke.edu/~tjl13/s101/slides/unit6lec3H.pdf) - can I derive the intervals from .632+ bootstrap method or is there a different way of getting these intervals?
> 
> Thank you!
> Raga
> 
> 
> On Wed, Mar 1, 2017 at 3:13 PM, Sebastian Raschka <se.raschka at gmail.com> wrote:
> Hi, Raga,
> I have a short section on this here (https://sebastianraschka.com/blog/2016/model-evaluation-selection-part2.html#the-bootstrap-method-and-empirical-confidence-intervals) if it helps.
> 
> Best,
> Sebastian
> 
> > On Mar 1, 2017, at 3:07 PM, Raga Markely <raga.markely at gmail.com> wrote:
> >
> > Hi everyone,
> >
> > I wonder if you could provide me with some suggestions on how to determine the confidence and prediction intervals of SVR? If you have suggestions for any machine learning algorithms in general, that would be fine too (doesn't have to be specific for SVR).
> >
> > So far, I have found:
> > 1. Bootstrap: http://stats.stackexchange.com/questions/183230/bootstrapping-confidence-interval-from-a-regression-prediction
> > 2. http://journals.plos.org/plosone/article/file?id=10.1371/journal.pone.0048723&type=printable
> > 3. ftp://ftp.esat.kuleuven.ac.be/sista/suykens/reports/10_156_v0.pdf
> >
> > But, I don't fully understand the details in #2 and #3 to the point that I can write a step by step code. If I use bootstrap method, I can get the confidence interval as follows?
> > a. Draw bootstrap sample of size n
> > b. Fit the SVR model (with hyperparameters chosen during model selection with grid search cv) to this bootstrap sample
> > c. Use this model to predict the output variable y* from input variable X*
> > d. Repeat step a-c for, for instance, 100 times
> > e. Order the 100 values of y*, and determine, for instance, the 10th percentile and 90th percentile (if we are looking for 0.8 confidence interval)
> > f. Repeat a-e for different values of X* to plot the prediction with confidence interval
> >
> > But, I don't know how to get the prediction interval from here.
> >
> > Thank you very much,
> > Raga
> > _______________________________________________
> > scikit-learn mailing list
> > scikit-learn at python.org
> > https://mail.python.org/mailman/listinfo/scikit-learn
> 
> _______________________________________________
> scikit-learn mailing list
> scikit-learn at python.org
> https://mail.python.org/mailman/listinfo/scikit-learn
> 
> _______________________________________________
> scikit-learn mailing list
> scikit-learn at python.org
> https://mail.python.org/mailman/listinfo/scikit-learn



More information about the scikit-learn mailing list