Feature Selection Using R | Machine Learning Models using Boruta Package

[@nkunam]

A true teacher who is not only selflessly sharing his knowledge, but also thanking all of us people who subscribe to learn from his teachings. Unlike most of the professionals in work places who are doing all sorts of silly things to protect their little turfs and safeguard the very little knowledge they have. What a difference! Thank you sir.

[@bkrai]

You are very welcome! 😊

[@jasonlee5814]

By far the best example on Boruta model on KZbin. Thank you Dr. Rai.

[@bkrai]

You are very welcome!

[@muhammadmurtala9045]

Your videos have been hugely instrumental for me learning R. I don't have the words that can adequately convey the magnitude of my gratitude to you, but let me follow the norm and say a BIG THANK YOU for your effort. Thanks a lot, sir!

[@bkrai]

Many thanks for your comments and feedback!

[@netmarketer77]

I learnt a lot from your videos Dr. Bharatendra. your helpful videos solve me a lot of problems and they are better than 3 hours lecture I attend in a class. Regards

[@bkrai]

Thanks for your positive comments!

[@sammy0722]

Thank you sir. I used to do these things in PYTHON programming and these things used to consume too much time. I find R- software to be a time saver. I enjoy R more than PYTHON.

[@bkrai]

Thanks for the feedback!

[@piuslutakome610]

Hell Prof. Rai, I am trying to select features with the Boruta package but getting errors. My data has 75 observations and 32 variables. The first two columns are character type with Column 1 having CowID's and column 2 has Days in Milk (DIM ie d-7,d+9, and d+21). When I tried running the Boruta, I got the error below, "1. run of importance source... Error in ranger::ranger(data = x, dependent.variable.name = "shadow.Boruta.decision", : Error: Unsupported type of dependent variable." What could be the problem?

[@aks1008]

Sir are there different feature selection functions for different algorithms in r or bortuta will work for all algorithms...like linear, logistic regression, support vector machine, cluster analysis etc..

[@bkrai]

For feature selection you can use a common method, but then develop your model using various methods using important features.

[@aks1008]

@@bkrai Sir and is there boruta () in python too...

[@dr.naeemhaider4747]

great video, I haven't seen any other channel with such detailed explanation with code. please keep making videos like these, you are such a good teacher. please make the video about time series forecasting and feature selection with the genetic algorithm.

[@bkrai]

Thanks for comments and suggestion, I've added them to my list.

[@db2885]

I an building a model for two parameters A and B . Do I need to select the explanatory variables that are common to both parameters A and B ? Thank you in advance.

[@bkrai]

Yes, A and B can be used in one column as independent variable. And then you will need data on explanatory variables for each row.

[@shivamnigam9628]

Sir, I have a task to predict energy rating transmission. But the problem is the dataset contains more than 70 variables and around 60 of them are categorical variables like energy transmission battery condition etc. Will it be wise to use boruta algo at the first hand , the data in several cols contains null and blank values ? And, how to convert 6 or more features into one feature?

[@bkrai]

Yes, boruta will help with feature selection. For reducing number of features, you can use PCA. I'm also including link for missing data. PCA: kzbin.info/www/bejne/haDaeH6EnMmiraM Missing data: kzbin.info/www/bejne/d5-an4OCf5WZqck

[@yashodharpathak189]

Thanks for the video. Can you please advice how to carry out feature selection when there is a mixure of continuous and categorical independent variables?

[@delt19]

Congratulations on reaching the 10k mark. Looking forward to many more of your videos.

[@bkrai]

Thanks!

[@ashok6644]

I've watched the video for "Feature Selection " which is commendable. Thank you for sharing your knowledge.

[@bkrai]

Thanks for comments and feedback!

[@tamtzeheuey]

Thank you Dr. Bharatendra Rai for providing very clear explanation of the feature selection (FS) in R. How can I modify a feature selection algorithm in order to enhance the performance of the FS algorithm?

[@bkrai]

As you saw in this video, it uses random forest and shadow variables. You can try any variation on your own.

[@lindanidube5714]

Sorry... mine keeps on giving me Error: Object “SelectedIndices” not found upon generating ROC curves. How do I deal with this issue, please?

[@bkrai]

which line in the video is causing this?

[@lindanidube5714]

@@bkrai thank you so much for your response... using your code, may you please also include ROC curves for Machine learning models. I was following your codes and they work perfectly fine. Also, please include a code to generate this on a single graph. Please🙏🏽 Machine learning algorithms: neural networks, decision tree, knn, SVM, random forest, Logistic regression , Naive Bayes and eXtreme gradient boosting. Your assistant will be greatly appreciated 😊

[@osowatejiri5930]

Thank you so much for this video. Please what packages can i install in R to add these libraries

[@bkrai]

Use these packages: library(Boruta) library(mlbench) library(caret) library(randomForest)

[@singhvaibhav033]

Sir, Amazing video one question though can we use boruta even when we have categorical variables in our data? If not, then please do feature selection video if there are mix of num & categorical variables in our dataset.

[@bkrai]

This will work fine with categorical independent variables.

[@singhvaibhav033]

Thank you Sir, I see not only your videos but all the comment replies as well (clears half my doubts from there), This is truly great of you, continue replying to important questions in comments. It helps all your students !

[@debasmitadey231]

If some variables of my data are non linearly correlated, will this package work ? or it will reject all those variables as their linear correlation is not significant ?

[@bkrai]

It takes care of non-linearity.

[@ramp2011]

Great video thank you. I have been using RandomForest varImp to do this. I am curious if Barota is a better package to identify the variable importance? Thank you..

[@bkrai]

It has several features that are very useful for handling situations with the lots of variables. For example, instead of typing names of 33 variables, you can get the formula easily and just copy/paste.

[@rameshsahni]

sir, I have a data of 140 band of wavelength with reflectance of a particular object, in that i want to select which band of wavelength is best suited to do the job. Is random forest model using boruta will work for the same?

[@bkrai]

It didn't fully understand your data, so difficult to suggest a method.

[@Gius3pp3K]

Love the videos. I have learnt so much from your KZbin channel. Thanks. Is there a way to use feature selection for a Neural network, for a large dataset, using Keras in R Studio please? I have tried to use LIME, with no joy.

[@bkrai]

Thanks for comments! I'll look into Keras question.

[@Gius3pp3K]

@@bkrai that would be greatly appreciated. Thanks

[@mustafacakir__]

Hi Mr. Rai. Firstly I would like to say thanks for your pure explanation of the topic. I tried the "boruta" function a few times with different seed numbers and got many different results for the important attributes. Is this function much sensitive to random seed numbers? Thanks.

[@bkrai]

Yes, it's because of the randomness.

[@abdulwaheedshaikh3745]

Sir, U r very excellent mentor for R. I love You for the sake of R. I am Ph.D. scholar from Chennai. Institute name is : Crescent Institute of Sci & Tech, Chennai.

[@bkrai]

Thanks for your comments!

[@tsumigoonetilleke4628]

Hi Bharatrendra, Thank you for the link. I watched the video, and it doesn't seem to have wrapper for the classification. Do you have any other videos with a wrapper? Thank you for your help. Cheers

[@MHRAJAI]

Thank you very much for your nice explanation. Can you please explain, what is the significance level to decide either variable is important or unimportant. If the Z score is significantly higher than the maximum Zscor of shadow features (MZSA) it is assigned as important feature. Can you pleas what is alpha here. Thank you 🙏

[@dhanashreedeshpande7100]

Wonderful video! From where are u searching such a different algorithms (such as this feature selection)? You can increase accuracy here further by using parameter tuning function in Random Forest. If possible please add video of web server log preprocessing also.

[@bkrai]

Thanks for the feedback and suggestion. I've added your suggestion to my list.

[@zaafirc369]

I love your videos. Thanks for the amazing work that you do. Do you offer online r courses? I really want to master machine learning using r programming

[@bkrai]

Thanks for your feedback! My online courses are limited to UMass-Dartmouth at this time. But you can learn many machine learning using r methods from this link: kzbin.info/aero/PL34t5iLfZddu8M0jd7pjSVUjvjBOBdYZ1

[@hanivlog774]

Great video and detailed information. Thumbs up for your hard work! Can we use boruta algorithm with other classification models such as SVM, NB, or MLP, etc?

[@ChungChingZhou]

Very clear articulation! thank you!

[@bkrai]

Thanks for comments!

[@deprofundis3293]

Thank you so much for this video; I just subscribed! But I have a question; you'd recommended this method for me on one of the multinomial videos (and I *cannot* thank you enough for the personalized feedback...you wouldn't believe how long I've been slogging through this on my own, desperate for more personalized guidance...). I was really impressed with the "boruta" performance; its tendency to keep ALL important variables actually makes the most sense for my data (but I fear that might not jive with regression). But the cross-validation is where things fall apart. I'd read that I shouldn't use RandomForest on too small of a dataset because there isn't enough data to partition, and my dataset is admittedly small. And when running this today, my accuracy during the final steps on the test dataset was awful.*** I was wondering: (1) Would you recommend an alternative that doesn't require partitioning of data, e.g., a LOOCV (leave-one-out cross-validation) regularization method, like glmnet? (2) Would it be valid for me to use the results of boruta to help inform model-building with AICc? Honestly, much of this machine learning is quite new to me. Small sample size issues are common in my field, but it seems like most people just use AICc to build their models, w/o external validation. I want to be rigorous enough to avoid Type I error, but I also need something sensitive enough to avoid Type II error on such a small dataset. ***It probably doesn't help that I have very low membership in a couple of the groups for my categorical response variable (I also plan to run it w/o them, but then my sample size is even smaller). I'm going to take a look at your video on class imbalance, but I'd be so grateful for more guidance about resolving my particular situation. I've been trying to find an appropriate approach for over a year now...

[@yousif_alyousifi]

Can Boruta deal with the missingness? which better to use all data or the train data? I

[@bkrai]

Yes, it should work fine. Also I used all the data as response was available for all rows. If the test data doesn't have response column, then only train data can be used.

[@sebismo]

Thanks for the video!!, Very usefull!. It works just for binary classification problems? What about for regressions problems?. Categorical features its also supported by this method?

[@bkrai]

Would work fine with categorical independent variables. Classification and regression would both work fine. For regression, when using test data, make sure you use root mean square error as confusion matrix is not valid there.

[@sudiptapaul2919]

Useful indeed! Lot's of struggle ends here. Thank you Professor.

[@bkrai]

You're very welcome!

[@ousmanelom6274]

Thank we can use this for categorical variable what test boruta use

[@bkrai]

You may refer to the following for more details: cran.r-project.org/web/packages/Boruta/Boruta.pdf

[@ravindarmadishetty736]

wonderful video sir....Please clarify my doubt if we are applying logistic regression can we avoid calculating WOE and IV to identify the strength of the attributes and considering them into the model? Also can we avoid data reduction technique when huge number of attributes?

[@bkrai]

Thanks for comments! What is WOE and IV that you referred to? If you would not like to remove any attributs where there are huge number of attributes, you can do PCA.

[@ravindarmadishetty736]

Thank you for replying sir....WOE stands Weight of evidence, IV stands Information value(like p-value criteria). We use them for continues variables and these splitting into bins. These are the keywords generally we use in credit scoring when applying logistic regression. These are useful in knowing the strength and identifying important variables in model. Now after looking this video....just doubted can we use directly proposed algorithm in this video instead calculating WOE and IV for important variables. These WOE and IV only considered for logistic regression and not for other classification models

[@bkrai]

Note that logistic regression is used only when your response variable is categorical. Usually we try more methods and look at confusion matrix and accuracy. Apart from logistic regression, you can also try random forest and see which method gives better result.

[@ravindarmadishetty736]

sir, give me your suggestion, i have 578 attributes and i have applied Boruta for featute selection. But, it is taking for 15 mins time for every importance running source..... So in this case, how to overcome this issue? I guess Factor or PCA can apply for data reduction Please advise

[@sumeet1509]

Thanks so much for this awesome tutorial. I am running Random Forests (RF) for academic research purposes. We really are not concerned in the first instance if we have a large number of features. We are more concerned about what to do with collinearity between some of the features (absolute r = 0.5 to 0.9). Some literature suggests that we can include correlated features in RF. What would you recommend? Can Boruta help with this issue? I notice that there is much collinearity in the Sonar dataset, especially amongst adjacent variables

[@bkrai]

With machine learning models such as RF, collinearity is not an issue.

[@vijaysrirambhatla3874]

Thanks for the explanation about Boruta, can you please provide a reference to get more information about the method.

[@bkrai]

Here is the link: cran.r-project.org/web/packages/Boruta/Boruta.pdf

[@joujoumilor2898]

thanks sir for sharing this amazing video with us, I have a question about the data should they be normalized or not ?

[@bkrai]

Normalizing has no negative effects. So I would say whenever there is a doubt, then definitely normalize.

[@joujoumilor2898]

thank you so much :

[@bemuzeeqtv]

Great video, I am following the steps but I would like to know how I can create my class for my dataset (if possible in R). Because it seems the Class was already defined in the dataset you used (M,R). I would appreciate any help. Thank you.

[@bkrai]

You can create a new column for that.

[@bemuzeeqtv]

@@bkrai Thank you. I did that

[@surbhiagrawal3951]

sir these boruta will work with any model where feature selection is required ?or is it mailnly with random forest ?.. I suppose we generally used it with classiifcation problem ?

[@bkrai]

Since random forest is used along with the idea of shadow attributes, it should work well in many situations. Also since random forest is used, should work both for classification and regression.

[@surbhiagrawal3951]

Also one more question , do the response variable needs to be categorical..? if it is continuous response variable , then also it will work ?

[@bkrai]

Yes

[@deepakbalajiselvam8067]

Great video with crystal clear explanation, many thanks

[@bkrai]

Thanks for comments!

[@kolozsie]

Thanks for the explanation. Tried to run the algorithm on radiomic features but it didn't found any important attributes ('No attributes deemed important'). How's that possible?

[@bkrai]

I used maxRuns of 500. You can increase it to a higher value if your data needs more runs.

[@melodicguitarist]

Very well explained Sir. Can you please let me know what set.seed () function do, and after watching your videos Can I practice by downloading the dataset from Kaggle ? Please let me know about this.

[@bkrai]

set.seed() helps to obtain repeatable results. Without this two people may run the same code but may get different results. I've provided link to data in the description area of the video. You can easily get the file there. Practicing with some Kaggle dataset is also a good idea.

[@anjaliacharya9506]

I install all packages mentioned in the video, I get result from print(boruta) but for plot(boruta) i get error message : Error in xy.coords(x, y, xlabel, ylabel, log) : 'x' is a list, but does not have components 'x' and 'y' what can i do, do i need some other packages?

[@bkrai]

Check earlier steps and also look at the structure of your data.

[@aks1008]

Dear Sir can we apply boruta () function for other algorithms e.g Linear Regression, Logistic Regression, Support Vector Machine etc along with Random Forest. Thanks

[@bkrai]

Once you select important features, the final model can be developed using other algorithms.

[@mondalsandip]

Hello sir, I have two datasets. One is the predictor matrix (species abundance data) and another one is response variables (Soil data) matrix. Can we use boruta in this type of situation? I want to see what are the soil factors influences my species community. I ran this in R, but showing problem like this: Error in ranger::ranger(data = x, dependent.variable.name = "shadow.Boruta.decision", : Error: Competing risks not supported yet. Use status=1 for events and status=0 for censoring. Thanks in advance

[@bkrai]

What was the code that you used?

[@mondalsandip]

@@bkrai Boruta( x, y, doTrace=2, maxRuns=100) Where x is the species abundance matrix Y is the soil data matrix

[@bharathjc4700]

Great Presentation sir. How is different from RFE recursive feature elimination?

[@bharathjc4700]

Good to see u sir.

[@bkrai]

Thanks!

[@bkrai]

In Boruta, importance significantly larger than those of shadow variables is used. In RFE, random forest with smallest error based on iterative removal of least important variables is used. Both methods are effective in handling feature selection.

[@bharathjc4700]

thank a ton for your valuable inputs sir.

[@abiani007]

How can I use this feature selection for regression? Shall I use the same technique as you have shown for regression purposes? Plz confirm.

[@bkrai]

With regression you can use statistical significance for deciding which independent variables to keep in the model.

[@abiani007]

Dr. Bharatendra Rai and sir plz share a video on FastICA

[@kellyng5474]

This video has been very helpful, however when running the confusion matrix i got an error as follows: > p = predict(rf70, test) > p > confusionMatrix(p, test$yVar) Error: `data` and `reference` should be factors with the same levels. All of my variables (x and y) are continuous variables. Any idea how to solve this issue? Thanks!

[@bkrai]

For response as continuous variable, you don't need confusion matrix. Probably you can try this: kzbin.info/www/bejne/lWTbfoaYfsmYaKs

[@kellyng5474]

@@bkrai Thank you very much! I will check out the video :)

[@bkrai]

You are welcome!

[@marcelofalchetti]

Your videos are really really useful and well explained, thanks for your work!

[@bkrai]

Thanks for the feedback!

[@kushxmen]

This video is great. Thank you so much. I have a question on Feature Tools (Python). Is there an equivalent package in r for doing the same? I want to create new features from the predictor variables I have before I perform Feature Selection.

[@bkrai]

I'm not aware of any equivalent package in r at this time.

[@kushxmen]

@@bkrai Thanks for the reply. I think as the R community we may need to develop one in future. Will embark on further research. Thanks again for the reply.

[@outinthebeach]

Thank you so much for your videos. You explain and articulate the steps so well by keeping it simple. It really helps me understand these models easily. Could you please help with or put up a complete lifecycle / steps right up to AUC for some common models like Random Forest, SVM, Logistic Regression...etc. that can be used as a template for model performance and model improvement. If possible. Thank you so much again.

[@bkrai]

Thanks for your feedback! You can find many of these methods at this link: kzbin.info/aero/PL34t5iLfZddu8M0jd7pjSVUjvjBOBdYZ1 I'll continue to add more.

[@jean-lucfanny4210]

Could you use the same Boruta algorithm with a numeric response instead of the class response (binary)? Since RF could be run using class response or numeric response. Please, guide me,

[@bkrai]

Thanks for the suggestion, I've added it to my list.

[@jean-lucfanny4210]

You could use Numeric as well. Your videos are life-changing. I really appreciate them.

[@bkrai]

Thanks!

[@faridehalasti3752]

many thanks, Boruta can cove Surv? how do it works? I need an Example.

[@playerjc9969]

Thanks very much for the awesome tutorial! It really helped me understand a bit of boruta. I have a question, I am currently running boruta and before doing tentative fix, i found three attributes that are tentative and when I did tentative fix, the other two got confirmed important and the other one is unimportant. The thing is, the unimportant attribute had higher normHits percentage than one of the confirmed attributes. The rejected attribute had 0.5306122 and the confirmed attribute had 0.4693878. Is this okay?

[@bkrai]

I would suggest run classification model with and without those variables and see if it helps to improve model performance or not. If performance improves keep it otherwise exclude.

[@RahilKhowaja]

Great. You made it very simple for us

[@bkrai]

Thanks for comments!

[@sumeet1509]

Thanks for the well explained tutorial. I used the algorithm on the same sample data set - Sonar. The results indicate v43,v44,v45,v46,v47,v48 and v49 as one set of confirmed features among many others. However, I also noticed that amongst these seven features there is much collinearity , with correlation coefficient 'r' ranging from 0.5 to 0.87. Is there any specification in the code that can be used to filter out highly correlated features ?

[@bkrai]

Note that random forest is not impacted by collinearity. However, if you were using regression, then it is definitely a problem and needs to be addressed.

[@devawratvidhate9093]

Gr8 Video sir Thanks Could you suggest any webscraping material in R ...From begginer to pro level

[@bkrai]

Thanks for the suggestion, I've added it to my list.

[@DnyaneshwarPanchaldsp]

can we extract features from text ......such as noun, verb, pronoun, etc ... for selecting features for aspect sentiment analysis

[@AliHoolash]

Thank you for this nice tutorial. In your example data set, all the variables (except the target variable) are numeric variables. Will Boruta also work on a mix of categorical and numeric variables?

[@bkrai]

Yes, it will work with both type of variables.

[@AliHoolash]

Thanks for your prompt reply.

[@bkrai]

welcome!

[@raghavendras5331]

No words to explain how informative and specific your videos are. great thanks for that. For ali's question to add in, we should do one-hot encoding before running Boruta on categorical variables?

[@bkrai]

Thanks for comments!

[@thourayaaouledmessaoud9223]

Thanks for this well- explained video. I just have one question could we use Boruta with clustering algorithms such as ( K-means, Dbscan ...) does it works??

[@bkrai]

Clustering is unsupervised learning method. There is no response variable.

[@kapilgupta8722]

Thank you Professor for such a nice explanation. I have a query that for a data where we have multicolinarity issue, Lasso removes the correlated variables, but Boruta doesn't remove them and it shows all the correlated variables as important. If this is the case then which one is more preferable? Can you please comment on such situation?

[@bkrai]

You can think of this as non-parametric as there are no statistical assumptions for this machine learning based method.

[@kapilgupta8722]

@@bkrai Thanks professor for the comment. If we want to show the comparison of logistic and Random forest. Then how will we comment about the variables and their significance? Or can you please suggest how to show the comparison between these two methods?

[@bkrai]

For comparison you can use confusion matrix, accuracy, etc. I would suggest you review this playlist for more: kzbin.info/www/bejne/oqakeneepMmEmrM

[@kapilgupta8722]

@@bkrai Thank you Professor :). WIll go through this lecture.

[@SaranathenArun11E214]

Sir, thanks, if we have 20 discrete variables, how to find the variable importance?

[@bkrai]

This method will work fine with discrete variables.

[@theahmads7590]

hello sir huge fan thanks for your effort, i tried installing Boruta in my r package but r says there is no such package,can you help me

[@bkrai]

Run these lines: install.packages('Boruta') library(Boruta)

[@abiani007]

Can you send some links for the ensembling of models output in R for regression purposes? Thanks in advance.

[@bkrai]

Here is the link: kzbin.info/www/bejne/nnSvfICfj6eHqLc

[@tsumigoonetilleke4628]

Hi, Thank you for your video. It's great. Do you have a video or any help document or training for implement a wrapper Naive Bayes Classifier? Can you help?

[@bkrai]

Try this: kzbin.info/www/bejne/iH3NhISamMxrd68

[@tsumigoonetilleke4628]

Thank you Bharatrendra. I'll try that. Thanks again

[@InfiniteSEOHallam]

Hi Bharatrendra, Thank you for the link. I watched the video, and it doesn't seem to have wrapper for the classification. Do you have any other videos with a wrapper? Than you for your help. Cheers

[@poojamahesh8594]

sir, i need to find the accuracy of the model after finding the variable importance...please tell me how to do sir..please

[@bkrai]

It's covered line-34 onward.

[@asifhayat4163]

Sir kindly make a vedio on boruta implemented on raster data Having dem and satelite imagery at the same time. Need ur help and waiting for your coding

[@bkrai]

Thanks for suggestion!

[@21bagong]

Dear Prof Rai, Would you mind explaining how boruta algorithm and function works?

[@bkrai]

It creates shadow variables for each independent variables and randomizes their values. Then the algorithm proceeds to check whether or not these variables exhibit better performance compared to their corresponding shadow variables.

[@21bagong]

@@bkrai Thank you Dr Rai

[@akshayagrawal7848]

I love your videos! Very helpful, got me through my deep learning grad class.

[@bkrai]

Thanks for comments and feedback!

[@SaranathenArun11E214]

happy to see you sir...thanks for all the videos and am great fan of you

[@bkrai]

I appreciate your feedback and support!

[@vijaysrirambhatla3874]

How can I see the complete list of important variables ? thanks

[@flamboyantperson5936]

Great Sir. First time ever I am seeing you on video. I am glad to see you Sir. Your video is video I have been waiting for a long time for your new videos. Thank You so much Sir.

[@bkrai]

Thanks for your comments !

[@Sandeep-sl7lp]

Sir do we need to standardise data or we can give non standardised data to boruta

[@bkrai]

It's not needed here.

[@Sandeep-sl7lp]

@@bkrai thanks a lot

[@bkrai]

You are welcome!

[@choubeyrajj]

After searching for a perfect example, my search ends here. Thanks for sharing this. While working on 1lac records, the process becomes really time taking. Is there any way to process it faster or any other method to apply this on large data sets.

[@bkrai]

Probably you can take a 30-40% sample and then apply feature selection to save time.

[@DeevyankarAgarwal]

Really it was a nice video, thank you Sir. It is taking too much time as my dataset contains 50000 rows and 57 columns. Can u suggest any algorithm for the same purpose that will take less time?

[@bkrai]

You can refer to this link and among 10 algorithms, xgboost is very quick: kzbin.info/aero/PL34t5iLfZddsQ0NzMFszGduj3jE8UFm4O

[@kamilchosta5526]

Amazing stuff. Thank you!

[@bkrai]

You are welcome!

[@sharjeelarain6897]

Sir your video is awesome please make videos on fb and KZbin analytics

[@bkrai]

Thanks for the suggestion! For KZbin and Twitter refer to: kzbin.info/www/bejne/ZqnWfmODl7eDfac

[@harunbakirci1781]

hi teacher first of all thank you very much your tutorials I have a little question variable importances are being diffrent in diffrent algorithms for example tv is better than radio in random forest algorithm but at xgboost algorithm radio is better than tv whichone is correct ? is it dependent that I use algorithm.

[@harunbakirci1781]

firstly I have used boruta. It gave same result with randomforest. But then I have used xgboost algorithm. It has given diffrent result from randomforest and neuralnet.

[@bkrai]

Yes, it can depend on the algorithm used and their rank on the list can change little bit.

[@harunbakirci1781]

@@bkrai okey my teacher thank you very much 👍👍👍👍

[@harunbakirci1781]

should we use boruta package or varImp or importance(randmforet(model)) or garson function in neuralnet which one should we use

[@bkrai]

Yes, that's ok as they are different methods.

[@deepakpanigrahi9601]

Can we also have some video (great learning tools) on Model deployment please

[@bkrai]

Thanks for the suggestion, I've added it to my list.

[@rajlaxmikati1175]

Actually my dataset has so many NAs when I am using boruta I am getting can't process NAs in input error how to handle that?

[@bkrai]

You can use this link for missing values: kzbin.info/www/bejne/d5-an4OCf5WZqck

[@rajlaxmikati1175]

@@bkrai thank you so much

[@bkrai]

welcome!

[@rajlaxmikati1175]

@@bkrai why working with very huge dataset soo I am getting error saying it can't process that much mb of data .. what should I do please suggest me some idea

[@bkrai]

In the 3rd RStudio window if you have too many big dataset, you can remove them by clicking on broom symbol. That will free up space.

[@syuhadaazamil4111]

i want to ask, if my datasets is to large, why when i run it takes longer times? hope you can reply as soon as possible to help me

[@bkrai]

That's natural. Larger dataset consume more computing resources that's why it takes more time.

[@syuhadaazamil4111]

@@bkrai Tq Dr. Do you think should or shouldn't i reduce the data in the dataset? because i want to make it faster.

[@syuhadaazamil4111]

@@bkrai actually i dont understand how to make a graph using boruta in R becuase my dataset is linear regression.

[@bkrai]

Sometimes when data size is too big, samples of sufficient size can be taken to reduce processing time.

[@bkrai]

It should still work as random forest method works for both categorical and numeric variable.

[@sm.melbaraj1682]

Can we use Boruta for data with categorical variables?

[@bkrai]

Yes

[@sm.melbaraj1682]

@@bkrai okay. Thankyou! Your lectures are very understandable since you explain it very simply.

[@bkrai]

Thanks for comments! My today’s live lecture will be at 7:30pm India time.

[@hans4223]

Simply awesome and excellent

[@bkrai]

Thanks for feedback and comments!

[@dennismontoro7312]

was your data already scaled or normalized before you run Boruta?

[@bkrai]

No, it was not scaled or normalized.

[@liangqueenie151]

Great video! Thank you Sir

[@bkrai]

Thanks for the feedback!

[@muralidhara2063]

Thank you for sharing your thoughts and congratulations..

[@bkrai]

Thanks for your comments!

[@Asdfasdffff]

Really like your videos 👍🏼. Please, could you made video about tenserflow in R? Something like image recognition, or working on GPU

[@bkrai]

Thanks for your feedback! You can find TensorFlow some of the videos here: kzbin.info/aero/PL34t5iLfZddtC6LqEfalIBhQGSZX77bOn

[@subaganesh552]

Sir, can you explain stepwise feature selection algorithm in r???

[@alipaloda9571]

Error: `data` and `reference` should be factors with the same levels. I got this error can you please help is that we have to convert every field as factor

[@bkrai]

Which line in the video are you referring to?

[@alipaloda9571]

@@bkrai I was trying to create confusion matrix sir At that point of the I got this error

[@bkrai]

I would suggest you check your data using str.

[@Ketonen08]

Thank you for making videos. How would you set up the Boruta-function if it was for regression? For example for a frequency model (poisson) or severity model (gamma). I added some code if it is easier to answer with an example. The code below is what I tried, but I don't know if it is correct. # Example install.packages("insuranceData") library("insuranceData") data(dataOhlsson) ds_in

[@bkrai]

Note that the algorithm automatically creates shadow attributes, you do not have to create them.

[@equbalmustafa]

Thankyou sir for your videos, which are helping us alot....

[@bkrai]

Thanks for your feedback!

[@poojamahesh8594]

fea

[@alamgirsarder1331]

Can Any one help about how to glmnet use for feature selection ?

[@bkrai]

Here is the link: kzbin.info/www/bejne/lWTbfoaYfsmYaKs

[@SandeepKumar-me6qr]

One of the Great video Sir which is really helpful.

[@bkrai]

Thanks for feedback!

[@Adityasharma-zb7no]

Sir can you please teach us Factor Analysis as well

[@bkrai]

Thanks for the suggestion! I've added it to my list.

[@poojamahesh8594]

fea

[@redarabie7098]

thank you Sir for this usefell video

[@bkrai]

Thanks for the comments!

[@thejll]

I do not understand how a variable van be less important than its shadow?

[@bkrai]

It can happen with a variable that doesn't contribute much to the model.

[@kevingeorgejohn9094]

Sir I have a doubt..does the run of importance source takes hours or even days to complete for (..., doTrace=2, maxRuns=500)?.. It's taking a very large amount of time for each run of importance source

[@bkrai]

It depends on how big your data is in addition to computing power.

[@kevingeorgejohn9094]

@@bkrai and if I give no maxRuns arguments? Will there be any default no of run of importance sources?

[@kevingeorgejohn9094]

@@bkrai after its execution..why does it show " no features deemed important and unimportant and all are tentative"?

[@bkrai]

It needs more runs

[@kevingeorgejohn9094]

@@bkrai how much would be needed minimum? Would 20 be enough? Or 15?