RandomizedSearchCV- Select the best hyperparameter for any Classification Model

[@ijeffking]

Hello Krish, with you learning the subject of Machine Learning is a breeze. The topics you have been choosing all along have been interesting and key to remaining focussed and motivated. Thank you so much.

[@krishnaik06]

Hey Jeff thanks for ur kind words. Please share with all your friends and tell them to subscribe the channel. It will motivate to upload more interesting content.

[@ijeffking]

Sure Krish. It goes without saying!!! Thank you again.

[@kumapawa]

Thank you so much for the video, your contribution to education is invaluable. Imma pray for you!

[@swathys7818]

Wow !! Very clearly explained Thank you so much

[@sachincw]

Hey Krish, Thanks for the video! You could have used the object instantiated by RandomizedSearchCV Class to make the predictions instead of instantiating the Randomforest class again with the best parameters. I am just wondering why you chose to instantiate the Randomforest class again.

[@himanshu8006]

great thanks for sharing the knowledge, it really helped

[@lucybennett2130]

you didn't use Xtrain Ytrain at the end. once you find params for x and y, then do you have to fit model for Xtrain, Ytrain again, right?

[@louerleseigneur4532]

Thanks Krish

[@michellelee7585]

so this or grid search would be better for nltk

[@adityarajora7219]

its like grid, whats random about that????

[@luisurena1770]

how would i go about it with a pipeline?

[@shrutijain7805]

Why have you used standard scaler here! We don't need it here, right?

[@haribattula5187]

thanks for the video krish!! actually I expected you to explain how randomizedcv works, does it perform on all combinations that can be made from passed dict like gridsearch or it takes some random combinations to avoid increase in computational power. If that is case then what if best hyper parameter values or ignored, is it a drawback of randomized search?....

[@futureofinformationalmedic2785]

Very nice overview

[@krishnaik06]

Thanks Tom please subscribe the channel for more interesting content

[@sachink110]

Hello Krish, Nice informative video. Can Randomized search CV be applied to decision tree/SVM/Logistic regression/NLP for checking hyperparameters? If yes, then the same parameter code needs to be put or different parameters should be put in that hyperparameter selection bracket?

[@explorenations892]

Sir how you took 40?

[@priyaprasad1124]

Sir can you please my doubt, if after applying randomizedsearchcv the best score is -3.99999 then what does it mean

[@ijeffking]

Krish, I have one question. While doing GridSearchCV you did not use ".values" while defining X&y.(you did not need an array). But here you have. Would you kindly explain why and when should one use ".values" in general?

[@krishnaik06]

Hey Jeff. I will use .values only when i need the dataset in array form. The reason i convert into arrays is that if we are using sklearn libraries it is always better to give in the array format

[@ijeffking]

Thank you so much. Added to my growing fund of knowledge.

[@amanjangid6375]

Why gridsearchcv taking more time-complexity than randomizedsearchcv

[@jitenkumarsahoo667]

Hi krish,I have a query like,when I use train test split it gives good accuracy.however,when I use cross val score,it gives inconsistent accuracy for some iterations(e.g 95,96,96,94,95,78,63,54,88,68).How will I deal this type of scenario?

[@prashanthpandu2829]

Hello, I have doubt that randomized serach_cv should be applied on train dataset or on whole dataset. Can u please reply me.

[@MasterofPlay7]

yes, but i think he cross validate the model at the end and the average score was 89 or so....

[@haribattula5187]

no we should apply only on train dataset.

[@sreelakshmigopi6078]

what if we do to predict if we have more than one classification with multiple hyper parameters

[@mohammadarif8057]

label encoding

[@mohammadarif8057]

Hello Sir their is one problem that again and again when i run the randomized search cv cell every time it comes with different best score and parameters in that case what is the solution

[@gujjalapatiraju7435]

use random_state=some integer in the model parameter. It will help you to select a particular random state and split's data in the same order in every run.

[@vaibhavchhabra6270]

Why is he standardizing when he is using Random Forest,which has no concept of eucledian distance?

[@RitwikDandriyal]

Hey, I too thought of the same. But then again, it could be a general approach. Because you don't decide a classifier right away when you're working on a dataset. So in the pre-processing its always better to scale your features because then during model selection, you can test multiple algorithms on your dataset be it tree based or euclidean distance based algorithms. If you're sure and have made up your mind from the very start that you'll be using a tree based approach, then yes, you don not have to scale. But again, I think here a general approach is shown.

[@harshgupta3641]

is it compulsory to do scaling for tree based algorithms?

[@gurjotsingh752]

no, as per the rule it is not advisable to do scaling in tree based algo( but no one is stopping you to do that.) it wont be useful anyway.

[@harshgupta3641]

is it compulsory to do scaling for tree based algorithms?

[@kadhirn4792]

Not required. Because decision tree makes decision based on the label and the tree traversal to a node will be same when you do with/without scaling.