I'm using 5 algorithms for machine learning analysis of my data, (coxph, coxtime, rfsrc, deephit, deepsurv)
My dataset contains 40 features and 1700 observations.
I used nested cross-validation (inner folds = 10, outer folds = 3) and benchmarking for optimizing hyperparameters and comparing algorithms accuracy.
Hyperparameters to be optimized included number of nodes of layers, epochs, dropout, alpha for neural network, and mtry, nodesize for rfsrc.
I used random_search method.
For termination rule, I've chosen 60 iteration.
I have started analysis at 9:00 am and now it is 2:00pm, but my computer is still analysing.
my OS is windows 10, CPU intel corei7 8550U 1.8-1.99 GHz, RAM 12GB.
I'm using R 4.0.3 and mlr3proba package.
Is it normal for machine learning analysis to last many hours?
Honestly I have read in literature that it may takes for hours, but I don't know does it apply for complexity of my analysis? or there is a problem ?

  • 2
    $\begingroup$ I advise you to log your program's progress. O/w no one can precisely answer your question. But, it seems you've 60 x 10 x 3 x 5 = 9K combinations to try out. Roughly, in ~5 hours, that means 30 combinations to try in a minute (one comb in 2 sec). Was that possible in your trials with your setup? $\endgroup$
    – gunes
    Apr 19, 2021 at 9:40
  • 2
    $\begingroup$ Thanks, According to your calculation and my setup experience, I think no, it is not possible to perform one comb in 2 second. so I think I should wait more. Because the same parameters analysis but with much lower iterations were all successful before. And also in the future I should log my progress before starting time demanding analysis. $\endgroup$ Apr 19, 2021 at 10:48


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Browse other questions tagged or ask your own question.