I like the logic of this video. You showed the baseline, then three additional methods, then compare them in the end. Thanks a lot for sharing the technique. The feature/target matrix is also very helpful. My question is the principle or concept behind the filter method, RFE, and boruta. Is it possible to do a video on them?
Please do more Data science-related content, It was very helpful I searched everywhere for feature selection videos and finally landed on this video and this was all I needed, the content is awesome and the explanation is as well!
Wow, this video is really helpful, a lot of interesting methods were shown. Thanks a lot. I like to ask you to make a future video covering how you perform feature engineering and model fine tuning 1:49
This is an incredibly helpful video. One thing I noticed is that all features are numerical. How do we approach feature selection with a mix of numerical and categorical features? Also, when we have categorical features, do we first convert them to numerical features or first do feature selection. A video on this would be really helpful. Thank you
You will need to convert the categorical features into numerical format by using label encoding which automatically converts it to numerical values or custom mapping where u can manually assign ur preferred values to the features. I hope it helps
I only have one book for now: Time Series Forecasting in Python. I'm writing another one right now: Time Series Forecasting Using Foundation Models (cool stuff!!)
in Variance threshold technique, if we use Standard scaler instead of Minmax scaler, the variance would be the same for all variables.... does it means we can eliminate this step and just use standars scaler?
I want to make LSTM time series, what should I do for this? I think the situation is different for time series. Would I be wrong if I use what you did? There is both trend and seasonality in the series.
Hi, when I use randomforest , DecisionTree and xgboost on RFE , even if all of them tree based models, they returned completely different orders. On my dataset has 13 columns, on xgboost one of feature importance rank is 1, same feature rank on Decisiontree is 10, an same feautre on Randomforest is 7. How can I trust wich feature is better than others in general purpose ? İf a feature is better predictive than others, shouldnt it be de same rank all tree based models ? I am so confused about this. Also its same on SquentialFeatureSelection
That's normal! Even though they are tree-based, they are not the same algorithm, so ranking will change. To decide on which is the best feature set, you simply have to predict on a test set and measure the performance to make a decision.
The dataset comes from the scikit-learn library! We are not reading a CSV file. As long as you have scikit-learn installed, you can get the same dataset! That's what we do in cell 3 of the notebook and it's also on GitHub!
hi, thanks for the content 🎉 i have a dataset of 62 features and 40k rows and im using RFE. but rfe takes so much time. 90 minutes passed but it is still working. is there a problem? is this normal? what do you think?
It's normal, as by default, it will use all features and remove them one at a time. You can increase the value of "step" or set it to a float between 0 and 1 to express a percentage. That controls how many features are removed at each iteration and should speed things up. So, if you set "step" to 5, then it removes 5 features at every iteration. If you set it to 0.05, then it removes 5% of all features at each iteration.
You can convert the label to numerical features by replacing them with numbers. If you have 3 labels in a feature, you could represent them with 0,1,2 there are different methods to use. Simpler one is .replace({})
I am a noob to data science and feature selection. Yours is the most succinct and clear lesson I have found... Thank you!
I like the logic of this video. You showed the baseline, then three additional methods, then compare them in the end. Thanks a lot for sharing the technique. The feature/target matrix is also very helpful.
My question is the principle or concept behind the filter method, RFE, and boruta. Is it possible to do a video on them?
Please do more Data science-related content, It was very helpful I searched everywhere for feature selection videos and finally landed on this video and this was all I needed, the content is awesome and the explanation is as well!
Hugely informative and educational content. Many feature engineering videos are not that instructive.
Really great content! Learnt a lot. Thanks for your hard work!
Congratulations and thanks for the excellent tutorial. You gained another subscriber👏👏👏
Amazing video and excelent didatic. Congrats for the great quality, helped me a lot!
It was great! Thanks for sharing your knowledge. Hope to see more of you.
This is extremely helpful and informative. Thanks a LOT!
Very interesting explanation and clear to understand. I was looking for this kind of tutorial. Subscribed👍
I am currently reading your book and it's amazing
what is the name of the book plz
Wow, this video is really helpful, a lot of interesting methods were shown. Thanks a lot.
I like to ask you to make a future video covering how you perform feature engineering and model fine tuning 1:49
Great explanation. Easy hands-on as well!!
Thank you!
Sensational video, thank you so much!
Thanks for this valuable work. Helps me learning the subject.
Marco's the man!
very helpful video and easy way to explain the content. thanks alot
Woah , much awaited 🎉 . Thanks for all the efforts put in sir . Looking forward to more such amazing content 🙂
This is an incredibly helpful video. One thing I noticed is that all features are numerical. How do we approach feature selection with a mix of numerical and categorical features? Also, when we have categorical features, do we first convert them to numerical features or first do feature selection. A video on this would be really helpful. Thank you
You will need to convert the categorical features into numerical format by using label encoding which automatically converts it to numerical values or custom mapping where u can manually assign ur preferred values to the features. I hope it helps
You will have to do the conversion before feature selection because machine learning models only learn from numerical data
Dear Marco Thank you.😀
Thank you! It's helpful!
Glad it helped!
Than you very much for the video
this is great! about how long did it take to do boruta for your dataset? like if i have 400 features and 1 million rows.. would that be impossible?
@@edzme possible, but it will take a while, for sure.
thank you for clean insides can you list out all your books
I only have one book for now: Time Series Forecasting in Python. I'm writing another one right now: Time Series Forecasting Using Foundation Models (cool stuff!!)
Thanks, good review
Awesome video
Thanks!
in Variance threshold technique, if we use Standard scaler instead of Minmax scaler, the variance would be the same for all variables.... does it means we can eliminate this step and just use standars scaler?
Thank you for sharing
I want to make LSTM time series, what should I do for this? I think the situation is different for time series. Would I be wrong if I use what you did? There is both trend and seasonality in the series.
interesting content much love it🥰
Hi, when I use randomforest , DecisionTree and xgboost on RFE , even if all of them tree based models, they returned completely different orders. On my dataset has 13 columns, on xgboost one of feature importance rank is 1, same feature rank on Decisiontree is 10, an same feautre on Randomforest is 7. How can I trust wich feature is better than others in general purpose ? İf a feature is better predictive than others, shouldnt it be de same rank all tree based models ? I am so confused about this. Also its same on SquentialFeatureSelection
That's normal! Even though they are tree-based, they are not the same algorithm, so ranking will change. To decide on which is the best feature set, you simply have to predict on a test set and measure the performance to make a decision.
Excellent video, however I'm preoccupied trying to figure out if having wine as a gas would make dinner parties better or worse. 🤔
i tried to search under your github aint get the data where is the data you work on?
The dataset comes from the scikit-learn library! We are not reading a CSV file. As long as you have scikit-learn installed, you can get the same dataset! That's what we do in cell 3 of the notebook and it's also on GitHub!
hi, thanks for the content 🎉 i have a dataset of 62 features and 40k rows and im using RFE. but rfe takes so much time. 90 minutes passed but it is still working. is there a problem? is this normal? what do you think?
It's normal, as by default, it will use all features and remove them one at a time. You can increase the value of "step" or set it to a float between 0 and 1 to express a percentage. That controls how many features are removed at each iteration and should speed things up. So, if you set "step" to 5, then it removes 5 features at every iteration. If you set it to 0.05, then it removes 5% of all features at each iteration.
@@datasciencewithmarco thanks man! i have fixed it. i used mutual information to decide the number of features and RFE to select the best features.
Can you teach how to do MRMR feature selection in ML?
pretty helpful!
I thought feature selection is done before model training. Am I wrong?
Yes correct
can you explain how we are performing feature selection for the multilabel problem?
You can convert the label to numerical features by replacing them with numbers. If you have 3 labels in a feature, you could represent them with 0,1,2 there are different methods to use. Simpler one is .replace({})
subscribed