1
$\begingroup$

Let's assume my dataset

D1 has the variables x1, x2, x3, x4, x5, x6, x7, x8, x9, x10

I have used Gradient boosting regression on this and want to score on a new data set having same set of variable(of-course) but the sequence are different.

D_scoring has the sequence like x1, x5, x8, x3, x4, x7, x9, x2, x6, x10

Does this difference in sequence will hamper my prediction(accuracy).

If yes, what is the reason behind that ?

P.S: The reason behind the question is i have done a GBM regression the socring was good with the Validation and CV but while scoring with new data the learning is very bad.

$\endgroup$
4
  • $\begingroup$ Do you mean if the order of columns will play any part in the predictions? $\endgroup$
    – LyzandeR
    Nov 19, 2015 at 16:08
  • $\begingroup$ If predicting new data is bad, it means your model isn't describing your phenomenon well enough. $\endgroup$ Nov 19, 2015 at 16:40
  • $\begingroup$ @LyzandeR : Yes you got it, I want to ask if the order of the columns does play any part in prediction. $\endgroup$
    – Amarjeet
    Nov 19, 2015 at 18:21
  • $\begingroup$ @RomanLuštrik : that might be the issue, but still want to explore any possibilities, as the model was working good in validation data and with cross validation too. Its the unseen data (at the time of scoring) where its showing very odd prediction. $\endgroup$
    – Amarjeet
    Nov 19, 2015 at 18:23

1 Answer 1

1
$\begingroup$

The order of the components (columns) of multivariate data almost always matters. Fix an order and stick with it.

$\endgroup$
2
  • $\begingroup$ you are savior, Thanks a lot my model works just great, but if possible can you provide me any link where i can get the information about the pitfalls. $\endgroup$
    – Amarjeet
    Nov 20, 2015 at 0:08
  • $\begingroup$ Just consider a simple example where you're trying to predict a student's exam score, $S$, from the hours per day spent working, $W$, and playing computer games, $P$. Perhaps on the test data, you learn a model like $S = 50 + 2W - P$: the longer you work and less you play, the higher your score. Now think what happens if you apply that model to students whose $W$ and $P$ data have been switched ... you'll end up predicting that the lazy game-players will score highest. $\endgroup$
    – Creosote
    Nov 20, 2015 at 22:30

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.