Introduction to building machine learning models in R with mikropml (CC124)

[@Riffomonas]

What questions do you have about using machine learning methods?

[@brindaraj5910]

If the model specificity ( 0.333) is not good for all the models then, how to improve the specificity? for instance: cv_metri~1 logLoss AUC prAUC Accur~2 Kappa F1 Sensi~3 Speci~4 Pos_P~5 Neg_P~6 1 0.938 0.167 0.889 0.549 0.944 0.372 0.971 0.980 0.333 0.962 0.5

[@Riffomonas]

@@brindaraj5910 Thanks for watching! There's no guarantee that you'll get a model that works well so you have to keep in mind that if there's no signal, then there's no modelling approach that will find it. That being said, you do seem to have a good AUC and sensitivity. There's a tradeoff between sensitivity and specificity so if you can reduce the sensitivity, you'll get a larger specificity

[@morgomi]

as a Turkish, I loved the package' name :d

[@Riffomonas]

Ha! Glad you like it 😂🇹🇷

[@truemusicmedia]

very informative! very well done lesson. Thank you very much.

[@Riffomonas]

Hey True Music - Thanks!

[@Riffomonas]

Hey True Music - Thanks!

[@mehrbodestaki6818]

Looking very forward to trying out mikropml! Perhaps this topic is covered in a future video I haven't seen yet, but it would be great if you could discuss dealing with highly imbalanced datasets (i.e. 450:50 control to case structure) when using these ML methods, or better yet show an example of how you typically deal with those types of datasets.

[@Riffomonas]

Thanks for the suggestion, maybe I'll do something like this in a future episode where we compare normal/adenoma to cancers. That would have about a 1/3 to 2/3 imbalance

[@yingdongli3433]

learning from your paper, hope the following video come soon.

[@Riffomonas]

Thanks! Look for more videos every Monday, Wednesday, and Friday at noon EST (US)

[@wapsyed]

Never thought I would watch Seth Rogen teaching ML in R 😅

[@YvonneVallesRodriguez]

Loving the videos, the energy and the help. I have tried to run mirkropml but i have a doubt. When looking at the number of samples in the results, it shows 41 samples with 28 predictors.... It does not make much sense as i had initially 203 samples. Is 41 referring to the samples used to test the model and not the training set? Thank you!!

[@8bitgerman477]

These videos are great. Thank you, I just had one question. What are the implications of getting a higher AUC and accuracy on the test data than on the training data?

[@Riffomonas]

Hmmm. That does seem weird. Is it a lot different? Maybe try a different modeling approach? Thanks for watching!

[@nguyenlephuong9489]

Hi, thank you for your introduction to mikropml. I am planning to implement the pipeline in my study. In your opinion, what is the minimum sample number to input into mikropml? Is n=20 samples would be enough? Or n should be around 100 samples?. Thank you once again for your time and our advice

[@Riffomonas]

Maybe for the logistic regression models that would be enough.

[@nguyenlephuong9489]

@@Riffomonas Thank you so much. I will try with n=20 first.

[@charlottebraley8702]

Hello, it is possible to use this package and do ML with 3 levels (values in your "srn")? Example, do we can have healthy, symptomatic, asymptomatic? Thank you very much

[@Riffomonas]

It is! Check out the documentation and send us a note if you can’t find it

[@mohammedarazzaq7847]

Thank you so much for the channel and amazing explanation Please, I want to ask you about "Adaptive neuro fuzzy inference system" which package we can use to implement it in R. Thank you so much

[@Riffomonas]

I did a google search for "CRAN Adaptive neuro fuzzy inference system" and it looks like there are at least three packages - FuzzyR, Anfis, and frbs. I'm not familiar with these but that's where I'd start

[@mohammedarazzaq7847]

@@Riffomonas Thank you so much for your kind help, really I much appreciated that. Thank you

[@bedece1549]

thank you so much for the video, very interesting package. But I have a doubt, is necessary separate a 20% validate (in the 80%) when use a cross validation method? thanks for you help

[@Riffomonas]

Yes. This is to make sure that you get proper hyperparameters and don't overfit the model

[@rishikeshdash12]

Sir, Please suggest any book for ML in R and ML on microbiome using R. I want to learn and understand parameters used in ML model at basic level.

[@Riffomonas]

Check out Julia Silge and Max Kuhns tidymodels book

[@rishikeshdash12]

@@Riffomonas Ok Sir, I will go through this book

[@rishikeshdash12]

Sir, I have one doubt when we were using machine learning or deep learning in Microbiome data for predicting healthy or diseased what type of normalization we should perform with otu counts ? should we prefer clr or relative abundance? sir, i have microbiome data and few clinical parameters (vitamin-d level, womac score, age, year of pain) as features or you can say it input variables to predict two output variable as healthy or diseased so what type of normalization i should prefer for meta data? shall we use scale() function for all the features or different normalization for above features? thank you sir!

[@Riffomonas]

For OTU data we subsample once to the same number of reads per sample. For those other factors it will depend on the type of data you have

[@rishikeshdash12]

@@Riffomonas for meta data I have (Age, Vitamin-d level, womac score, year of pain), Shall I use Scale() function for this, Because it is similar to Pre_process function in mikropml?

[@rishikeshdash12]

@@Riffomonas Sir in last video (CC126) you have taken relative abundance then you again normalize them by scale and center? Reason?

[@rishikeshdash12]

what is fit_result here?

[@Riffomonas]

It’s the amount of blood in a stool sample

[@rishikeshdash12]

@@Riffomonas ok sir, thank you sir

[@nosaosawe3158]

What are your social media handles sir? I really love your works.

[@Riffomonas]

Thanks! @PatSchloss on twitter