"Brute force feature selection"

ammarghammargh Member Posts: 27 Maven
edited June 2019 in Help
Shouldn't brute force feature selection return the best performance? The performance I had using features I have selected manually was better than the performance using features returned from brute force selection component.

Is this normal?

Answers

  • MariusHelfMariusHelf RapidMiner Certified Expert, Member Posts: 1,869 Unicorn
    Hi,

    how big is the difference between your manual performance and the performance of the features found by brute force? Are you using a Cross Validation?

    Please keep in mind that by default the X-Validation always uses random splits, and thus small performance changes can be produced by randomness. To enforce the same splits in all X-Validations in all iterations, and also in your manual evaluation, you can set the local random seed of all X-Validation operators to a constant. Then only the performance of the algorithms and the features is compared, and the factor "random" is eliminated.

    Best regards,
    Marius
  • ammarghammargh Member Posts: 27 Maven
    I see your point.
    I will follow your advise
    Thank you.

  • mafern76mafern76 Member Posts: 45 Contributor II
    It should be impossible for a manual selection to do better than a brute force selection, because the latter simply tries all possible combinations.

    I agree with Marius, your results are biased because of randomness and I might add maybe due to high variation in your model algorithm.
Sign In or Register to comment.