I think instead of dropping "either of" 2 highly correlated features, we should check from both of them how each of them correlates with the target as well and then drop the less correlated with the target variable. Which might increase some accuracy instead of considering dropping whichever comes first. Again, I think it is.
@@niveditawagh8171 you only drop when two feature variables are highly correlated but you don't have to drop feature that is less correlated with target variable because less correlated feature with target variable could be a good predictor variable in combination with other features.
If you are transporting ice-cream in a vehicle, the number of ice-cream sticks that reach the destination is inversely proportional to temperature, higher the temperature, lesser are the sticks. If you want to effectively model the temperature of the vehicle's cooler and make it optimal, you need to consider this negatively correlated features, outside air temperature and number of ice-cream sticks at the destination.
Sir, could you please upload more videos on feature selection to this playlist? It is very amazing. I followed all the videos from feature engineering playlist. You are doing a great work. Thank you.🙏🏻
In which order should u do the feature selection steps? 0. Clean the dataset, get rid of NaN and junk values. Check format for datatypes in testset etc 1. Use z-method to eliminate outliers 2. Normalize the train_X data 3. Check correlation between x_train variables and y_train. Drop variables that have a low correlation with the target variable. 4. Use pearsons correlation test to drop highly correlated variables from x_test 5. Use variance threshold method to drop x_train variables with low variance. All variables that have been removed from the x_train data should be removed from the x_test aswell. 6. Fit x_train and y_ train to a classification model 7. Predict y(x_test) 8. Compare the predicted y(x_test) output with y_test to calculate accuracy 9. Try different classification models and see which one performs the best (have the highest accuracy) Is this the right order? Have I missed something?
I want to point out a veryyy important concept which is missing in this video discussion: Suppose 2 input features are highly correlated then it's not like that , I can drop any between those 2 , then I have to check which feature between those 2 has weaker correlation with output variable , that one has to be dropped.
Let's say variables x, y and z are all strongly correlated to each other. You would only need to use one of them as a feature. By saying [df.corr()>0.7 or df.corr()
Thanks krish, You've earned a rocket point from me :) Would have been nice, if the function also printed which feature it was strongly correlated with: because from the code you dropped all the features that met the threshold, not one was kept.
since we are giving only one positive value for threshold, the code abs allows check for both negative and positve values with threshold, so i feel its better if it stays
In this video it's said negatively correlated features are both imp. lets take an example, when we have both percentage and ranks in a dataset, for 100% we have 1 in rank and 60% lets say 45(last) in rank. both resemble the same importance in the dataset. So what I think is we can remove one feature among those 2 features, otherwise we will be giving double weightage for that particular feature. Hope someone can correct this if I was wrong.
I have a doubt. Suppose if A and B have correlation greater than threshhold and the loop includes column A from the pair. Further B and C are highly correlated(although C is not highly correlated with A)and the loop includes B in the list. Now if we drop A and B wouldn't that affect the model as both A and B will be dropped?
We cheak the correlation between inputs and the output so why you drop output column and then cheak correlation we use a VIF (variance inflection factor) to cheak the relationship between inputs and the threshold value is preffer 4.
Hi friend, I think the correlation function is removing more than expected because when the fors loops are iterating not validate if for a value > threshold the column and index already was removed before. I corrected the function and in this case the features removed are these: {'DIS', 'NOX', 'TAX'}. Also I tested creating the correlation matrix again and verify that there is not values > threshold. Please can you check it. def correlation(dataset, threshold): col_corr = set() corr_matrix = dataset.corr() for i in range(len(corr_matrix.columns)): for j in range(i): if abs(corr_matrix.iloc[i, j]) > threshold: if (corr_matrix.columns[i] not in col_corr) and (corr_matrix.index.tolist()[j] not in col_corr): colname = corr_matrix.columns[i] col_corr.add(colname) return col_corr
Wonderful explanantion. Krish as mentioned in video you said you upload 5-6 videos for feature selection. Can you please share the link for rest of them.
Hi Krish, I checked it somewhere and I think if the dataset has perfectly positive or negative attributes then in either case there is a high chance that the performance of the model will be impacted by Multicollinearity.
Great tutorial, but I think you're mistaken about the abs(). You're actually considering both with abs(). If you remove abs() and you keep the > inequality then a 0.95 would be > Thresh=0.9, but -0.99 would not satisfy this condition! If you want to remove abs(), then you need to test 2 conditions, like if corr_matrix.iloc[i,j] > +1*thesh (assuming thres is always +ve) and corr_matrix.iloc[i,j]
Hello nice video, how to do feature selection if we have more than one target variable? i.e. In case of MultiOutput Regression problem how we can do feature selection. do we have to perform the pearson correlation individually on each of target variable or is there another convenient way that can solve the problem?
Pearson's correlation only works with numeric features. However, if you want to explore the categorical features, you can use Pearson's Chi-square test. You can use the SKBest from scikit-learn and chi2. Hope it helps!
Sir, what you've shown in the last of this video, in that big data project, after deleting those 193 features, how I can deploy the model? Please share a video (or link if you have in your playlist) the deployment phase after deleting features. Thanks. ❤
The function used in the example will not deliver high correlation with the dependent variable. Because at the end you dropped the columns without being checking the correlation with dependent variable.
Two quick questions: (1) Why not remove redundant features, ie highly correlated variables, from X before splitting it into training and test? What would be wrong with this approach? (2) If one features variable is correlated with a value of 1 and another variable with a value of -1 with regard to a given feature, are these also considered redundant?
Hi Krish while removing the correlated features we haven't checked that the independent variable is corelated to dependent variable. As you said in staring we should not remove the features that are highly correlated to dependent variables so while generating the heatmap should we include the dependent variable also ? let me know if my understanding is correct?
Hi Ankit, If we include the dependent variable in this feature selection process, the accuracy of our model might get compromised. Also if you can see in video if 2 features are highly correlated we are only removing 1 feature. So if that feature has good correlation with dependent variable which we don't know yet it is still in the dataset. (As we have dropped only one feature out of those 2)
Why we are droping highly correlated feature after spliting train and test either it is easy to drop features from original data set and then we can simply split the dataset?❓😕🤔
If you're a student and have time to explore, please go ahead and implement it from scratch. It'll really help you to not only understand the basic working but also the software development aspect of creating any model (refer sklearn documentation and source code) and get to know more about industry level coding practices.
Hi sir, there's an obvious flaw in this approach. You can't drop all correlated features, but only some of them. e.g. perimeter_mean & area_se are highly correlated (0.986507), and they both appear in your corr_features. However, you can't drop all of them because from pairplot, you could see perimeter_mean has a clear impact on the test result.
Dear teacher, May I ask a question? In my case, I want to predict sale of 4 products with weather forecast information, season and public holiday one week ahead. So, do I need to organize weekly based data? When we use SPSS, we need to organize weekly data, how about Machine Learning? I feel confused for that. In my understanding, ML will train the data with respect to weather information. So, we don't need to organize weekly data because we don't use time series data. Is it correct? Please kindly give me a comment.
What do you think about feature reduction using PCA, looking for a correlation between each feature and principal components, and then using those who have the most number of correlation that is great than 50% (or any other)?
If absolute is not used then threshold can not be 0.85. If any features are highly co-related negatively like -0.85 Still it wont qualify for the drop. Hence Absolute is necessary.
Hi kris, in multicollinearity conceps we have both corrlation matrix as well as VIF to remove the collinearity. Which method is best or does that depend upon data
@@krishnaik06 i worked on a dataset which was highly correlated features and both these methods gave me different results. Hence was confused which method to use. Thats why this question. Thanks
Pearson correlation coefficient only measures the linear relationship between features. This approach may not be effective if there are non-linear relationships between features.
Instead of doing X_train , x_test split, if we find correlation of the whole data and then we compare correlated column's correlation with the dependent feature and then drop only those features among the correlated columns which are less correlated?....does my question makes sense? if it does, would it affect the model?
I believe those should be two separate questions. Regarding the split, it is necessary to split before getting correlation to understand its effect on the test data. If you do not split, then when testing, you're already assuming the correlation to be present in the test data and thus overfitting. Remember, the actual "test" data will always be unknown to us, and the split helps us validate the model and generalize it for the future unknown data. For the second question: Yes, that makes sense to me. After getting the "multi-correlated" columns, we can calc the correlation of each with the target, and drop the ones with low absolute correlation.
@@PraveenKumar-pd9sx if we check correlation of whole data rather than splitting(X_train, X_test). there is a chance that the correlation of whole data will be slightly different than the correlation if we had split. this might give us a better result on the validation (X_test) but would not perform on the actual test data when we deploy it in real world. this is my understanding from @Y S's comment.
Hi everyone i need one help. this technique to select numerical features only. Suppose we have done one hot encoding on categorigal data and converted into numerical then can we apply this technique on that features as well(entire data set with numerical column and categorical column converted into numerical with some encoding technique.) Kindly help me to understand.
Hello Sir my dataset contains 17000 features, when I execute corr() its taking more than 5 minutes to execute and also for generating heatmap memory related error generating. Can you help to solve the issue?
Being in a teaching profession ,I assure this is the best explanation about Pearson correlation.. Please make more likes.
I think instead of dropping "either of" 2 highly correlated features, we should check from both of them how each of them correlates with the target as well and then drop the less correlated with the target variable. Which might increase some accuracy instead of considering dropping whichever comes first. Again, I think it is.
good point
you can check importance value of each using RF and one can be dropped which has less importance value
Good point
Can you please tell me how to drop the less correlated variable with the target variable?
@@niveditawagh8171 you only drop when two feature variables are highly correlated but you don't have to drop feature that is less correlated with target variable because less correlated feature with target variable could be a good predictor variable in combination with other features.
If you are transporting ice-cream in a vehicle, the number of ice-cream sticks that reach the destination is inversely proportional to temperature, higher the temperature, lesser are the sticks.
If you want to effectively model the temperature of the vehicle's cooler and make it optimal, you need to consider this negatively correlated features, outside air temperature and number of ice-cream sticks at the destination.
Very comprehensive explanation for someone from non AI background. Thanks Sir keep up the good work!
Sir your channel is a perfect combination of sentdex and statquest. You are doing a great work 🙌more power to you!!
Sir, could you please upload more videos on feature selection to this playlist?
It is very amazing. I followed all the videos from feature engineering playlist. You are doing a great work.
Thank you.🙏🏻
Any word is not sufficient to thank you for your work sir ....🙏🙏
In which order should u do the feature selection steps?
0. Clean the dataset, get rid of NaN and junk values. Check format for datatypes in testset etc
1. Use z-method to eliminate outliers
2. Normalize the train_X data
3. Check correlation between x_train variables and y_train. Drop variables that have a low correlation with the target variable.
4. Use pearsons correlation test to drop highly correlated variables from x_test
5. Use variance threshold method to drop x_train variables with low variance.
All variables that have been removed from the x_train data should be removed from the x_test aswell.
6. Fit x_train and y_ train to a classification model
7. Predict y(x_test)
8. Compare the predicted y(x_test) output with y_test to calculate accuracy
9. Try different classification models and see which one performs the best (have the highest accuracy)
Is this the right order? Have I missed something?
I want to point out a veryyy important concept which is missing in this video discussion:
Suppose 2 input features are highly correlated then it's not like that , I can drop any between those 2 , then I have to check which feature between those 2 has weaker correlation with output variable , that one has to be dropped.
what do you mean by weaker? do you mean the most negative?
@@siddharthdedhia11, here , weaker means lesser correlation with output feature .
@@KnowledgeAmplifier1 so for example between -0.005 and -0.5 , -0.005 is the one with lesser correlation right?
@@siddharthdedhia11 yes , correct as correlation value towards 0 is considered as less value and towards 1 or -1 means strong relationship :-)
@jayesh naidu
I think the abs is important since it's like having two rows one being the opposite of the other
Yes, I think so too. If changes to one feature affects another feature, they are dependent, in other words, they are correlated.
Sir, the videos you uploaded on feature selection helped a lot ! , Please upload the rest tutorials and methods too! Eagerly waiting for it !
I write the threshold code simply like [df.corr()>0.7 OR df.corr()
Let's say variables x, y and z are all strongly correlated to each other. You would only need to use one of them as a feature. By saying [df.corr()>0.7 or df.corr()
amazing teaching skills you have bhaai ... THNX
thank you sOOo much , perfect explaining :) good luck with your channel that is recomended
Your knowledge is really invaluable. Thanks
I think it all depends on domain that whether to involve the neg corr or not , or we can train two diff models and compare their scores , Thanks Sir
Thanks krish,
You've earned a rocket point from me
:)
Would have been nice, if the function also printed which feature it was strongly correlated with:
because from the code you dropped all the features that met the threshold, not one was kept.
GREAT CONTRIBUTION SIR.... THIS CHENNAL SHOULD 20M SUBSCRIBER🤘🤘
since we are giving only one positive value for threshold, the code abs allows check for both negative and positve values with threshold, so i feel its better if it stays
This was incredibly helpful; thank you for the great content!
Well explained. Really great work sir. Thank you very much
In this video it's said negatively correlated features are both imp. lets take an example, when we have both percentage and ranks in a dataset, for 100% we have 1 in rank and 60% lets say 45(last) in rank. both resemble the same importance in the dataset. So what I think is we can remove one feature among those 2 features, otherwise we will be giving double weightage for that particular feature. Hope someone can correct this if I was wrong.
watching this video from Boston (BU Student
Thanks a lot for very clear explanation.❤
Sir make video on how to select features for clustering?
I have a doubt. Suppose if A and B have correlation greater than threshhold and the loop includes column A from the pair. Further B and C are highly correlated(although C is not highly correlated with A)and the loop includes B in the list. Now if we drop A and B wouldn't that affect the model as both A and B will be dropped?
again I wish if you explain how to handle the test set...but the explination is excellent am really gratful
Great tutorial it helps a lot thanks @Krish Sir
great video. very informative and educative. Thank you
We cheak the correlation between inputs and the output so why you drop output column and then cheak correlation we use a VIF (variance inflection factor) to cheak the relationship between inputs and the threshold value is preffer 4.
Hi friend, I think the correlation function is removing more than expected because when the fors loops are iterating not validate if for a value > threshold the column and index already was removed before. I corrected the function and in this case the features removed are these: {'DIS', 'NOX', 'TAX'}. Also I tested creating the correlation matrix again and verify that there is not values > threshold. Please can you check it.
def correlation(dataset, threshold):
col_corr = set()
corr_matrix = dataset.corr()
for i in range(len(corr_matrix.columns)):
for j in range(i):
if abs(corr_matrix.iloc[i, j]) > threshold:
if (corr_matrix.columns[i] not in col_corr) and (corr_matrix.index.tolist()[j] not in col_corr):
colname = corr_matrix.columns[i]
col_corr.add(colname)
return col_corr
Hi krish please a make a video on complete logistic regression for Interview preparation
Great video!! Thank you!👍👍💖
thank you, so usefull, good teacher
Nice! please upload more on this topic!! thank you!
Wonderful explanantion. Krish as mentioned in video you said you upload 5-6 videos for feature selection. Can you please share the link for rest of them.
Very helpful . Thank you sir.
Thank you man, good for my assignment
Hi Krish,
I checked it somewhere and I think if the dataset has perfectly positive or negative attributes then in either case there is a high chance that the performance of the model will be impacted by Multicollinearity.
Thanks sir for the good job you have been doing . God bless you. Please sir my question is can we use correlation on image data. Thanks
Nice explanation.
Thanks so much! very useful. you are so good
Sir , can you please tell which website should I refer if I want to start reading white papers.... Please please do reply....I follow all ur videos!!
waiting for more videos in the playlist
You are a legend!!🤘🤘
The abs takes care of both positive and negative numbers. If not specified, the function will only take care o positively correlated features
please clear it the below
if any independent variables are highly corelated we shouldn't remove them right because its give very positive outcome
Great tutorial, but I think you're mistaken about the abs(). You're actually considering both with abs(). If you remove abs() and you keep the > inequality then a 0.95 would be > Thresh=0.9, but -0.99 would not satisfy this condition! If you want to remove abs(), then you need to test 2 conditions, like if corr_matrix.iloc[i,j] > +1*thesh (assuming thres is always +ve) and corr_matrix.iloc[i,j]
Hello nice video, how to do feature selection if we have more than one target variable? i.e. In case of MultiOutput Regression problem how we can do feature selection. do we have to perform the pearson correlation individually on each of target variable or is there another convenient way that can solve the problem?
wonderful tutorial sir!!
Thanks Krish 😊
General Question - What if we drop few of the import features from and data and train again ? Will the accuracy drop ? or precision ?
You are doing a great job but can u please do similar series on categorical features in a regression problem?
Hi, thanks for the lecture. What if we have a dataset in which categorical and numeric features are present. Will pearson's correlation be applicable?
Pearson's correlation only works with numeric features. However, if you want to explore the categorical features, you can use Pearson's Chi-square test. You can use the SKBest from scikit-learn and chi2. Hope it helps!
Perfect defence on 3rd place!
Should small values of correlation such as -0.95 be deleted or they are good to train our model and they should stay in data frame?
Why you don't use corr_features = correlation( X , 0.7 ) instead of X_train. (Please look at 08:22)
Sir, what you've shown in the last of this video, in that big data project, after deleting those 193 features, how I can deploy the model? Please share a video (or link if you have in your playlist) the deployment phase after deleting features. Thanks. ❤
Thank you Krish,
Hello Krishna thanks for your video but along with please explain real life use as well. Where can we use in real life.
LOL, you are jsut amazing Boss
If idea is to remove highly correlated features, then both highly positive and negative correlation should be considered!!
Thank you for such a nice explanation. Does having 'abs' preserve the negative correlation ??
The function used in the example will not deliver high correlation with the dependent variable. Because at the end you dropped the columns without being checking the correlation with dependent variable.
Two quick questions:
(1) Why not remove redundant features, ie highly correlated variables, from X before splitting it into training and test? What would be wrong with this approach?
(2) If one features variable is correlated with a value of 1 and another variable with a value of -1 with regard to a given feature, are these also considered redundant?
Great explanation :)
Amazing content!~
Hi Krish
while removing the correlated features we haven't checked that the independent variable is corelated to dependent variable. As you said in staring we should not remove the features that are highly correlated to dependent variables
so while generating the heatmap should we include the dependent variable also ?
let me know if my understanding is correct?
Hi Ankit,
If we include the dependent variable in this feature selection process, the accuracy of our model might get compromised. Also if you can see in video if 2 features are highly correlated we are only removing 1 feature. So if that feature has good correlation with dependent variable which we don't know yet it is still in the dataset. (As we have dropped only one feature out of those 2)
Why we are droping highly correlated feature after spliting train and test either it is easy to drop features from original data set and then we can simply split the dataset?❓😕🤔
Sir, kindly make a video on embedded methods of feature selection..
Thanks, Sir.
Should i practice by writing my own full code including the hypothesis functions, cost functions, gradient descent or fully use sklearn?
If you're a student and have time to explore, please go ahead and implement it from scratch. It'll really help you to not only understand the basic working but also the software development aspect of creating any model (refer sklearn documentation and source code) and get to know more about industry level coding practices.
Can we drop features while comparing correlation of dependent variable with independent variables by taking some threshold....!
Hi sir, there's an obvious flaw in this approach. You can't drop all correlated features, but only some of them. e.g. perimeter_mean & area_se are highly correlated (0.986507), and they both appear in your corr_features. However, you can't drop all of them because from pairplot, you could see perimeter_mean has a clear impact on the test result.
Dear teacher, May I ask a question? In my case, I want to predict sale of 4 products with weather forecast information, season and public holiday one week ahead. So, do I need to organize weekly based data? When we use SPSS, we need to organize weekly data, how about Machine Learning? I feel confused for that. In my understanding, ML will train the data with respect to weather information. So, we don't need to organize weekly data because we don't use time series data. Is it correct? Please kindly give me a comment.
Another great video!!!
What if we have some features numerical and some features are categorical against categorical output .. which feature section method will be helpful
Hi Krish.. how about using an VIF to find the correlated features?
What do you think about feature reduction using PCA, looking for a correlation between each feature and principal components, and then using those who have the most number of correlation that is great than 50% (or any other)?
Sir when you will Upload Next video of this playlist (Feature Selection)
Hello can I ask a question ? Is Pearson Correlation the same as Correlation-based Feature Selection ?
Best!
Of the highly correlated columns, Should we not keep one of the columns in our X_train dataset?
what is the importance of random_state in train_test split ? How the values of random_state (0,42,100 etc.) affect the estiamation???
Krish, can we not use VIF for collinearity?
If absolute is not used then threshold can not be 0.85. If any features are highly co-related negatively like -0.85
Still it wont qualify for the drop. Hence Absolute is necessary.
Do we need use the entire datasets for correlation testing. Are we not missing something by considering the train set only?
Hi Krish,how to check in case of categorical variables
Hi kris, in multicollinearity conceps we have both corrlation matrix as well as VIF to remove the collinearity. Which method is best or does that depend upon data
Both are good...u can use any of them
@@krishnaik06 i worked on a dataset which was highly correlated features and both these methods gave me different results. Hence was confused which method to use. Thats why this question. Thanks
But I have vif was much more good
Hi. megala.. What is VIF. Can you pls tell me
@@PraveenKumar-pd9sx in short VIF is Variation inflation factor which also helps in finding multicolinearity between independent variables.
could you please make a video on " how auto encoder can be used to extract importance feature " ?
If I have 3 correlated columns, I should drop 2 out of 3 right ? why do you drop all correlated features from training and testing set ?
Is there any way to apply this approach to a classification problem where the target variable is 1 or 0?
Pearson correlation coefficient only measures the linear relationship between features. This approach may not be effective if there are non-linear relationships between features.
Instead of doing X_train , x_test split, if we find correlation of the whole data and then we compare correlated column's correlation with the dependent feature and then drop only those features among the correlated columns which are less correlated?....does my question makes sense? if it does, would it affect the model?
Same doubt
I believe those should be two separate questions.
Regarding the split, it is necessary to split before getting correlation to understand its effect on the test data. If you do not split, then when testing, you're already assuming the correlation to be present in the test data and thus overfitting.
Remember, the actual "test" data will always be unknown to us, and the split helps us validate the model and generalize it for the future unknown data.
For the second question:
Yes, that makes sense to me.
After getting the "multi-correlated" columns, we can calc the correlation of each with the target, and drop the ones with low absolute correlation.
@@YS-nc4xu Why should we split before the correlation check
@@YS-nc4xu YES!! i get it now, Thank you for sorting the issue!
@@PraveenKumar-pd9sx if we check correlation of whole data rather than splitting(X_train, X_test). there is a chance that the correlation of whole data will be slightly different than the correlation if we had split. this might give us a better result on the validation (X_test) but would not perform on the actual test data when we deploy it in real world.
this is my understanding from @Y S's comment.
Multi collinearity has checked but what about the Correlation of dependent vs independent variables
How to check correlation between numerical column (input) and categorical output(in the form of 0s and 1s)
How do you handle correlation for Categorical variables?
Hi everyone i need one help.
this technique to select numerical features only.
Suppose we have done one hot encoding on categorigal data and converted into numerical then can we apply this technique on that features as well(entire data set with numerical column and categorical column converted into numerical with some encoding technique.)
Kindly help me to understand.
Hello Sir my dataset contains 17000 features, when I execute corr() its taking more than 5 minutes to execute and also for generating heatmap memory related error generating. Can you help to solve the issue?
sir when the remaining videos on features selection are coming