Great question! Until I do a topic on Factor Analysis Regression (FCR), you can take a look at this video: th-cam.com/video/H45NWCzIDkY/w-d-xo.html (I go over the logic on how to use the transforms for other modeling procedures -- its the same process)
How do you know the # of factors that you look for? Is this based from your research context, or can you read the is from the date by experimenting with more or less factors? & what if you have contrasting factor loadings, can you invert these? (I.E.: I have question on how challenging something is, and after how easy it is which is in contrast)
I arbitrarily stated that the number of factors be 3 in this video. However, a more "scientific" way of determining number of factors is either 1) using a scree plot (elbow method) or you can check out this paper here: ruscio.pages.tcnj.edu/files/2016/08/Ruscio-Roche-2012-PA-Factor-Analysis.pdf on a more mathy way of finding out number of factors to optimize for. (factor loadings) You can also attempt to gridsearch for the number of factors (brute force based on model results) I'm not sure what you mean by contrasting loading factors?
@@SpencerPaoHere Thank you for your response and the source. By contrasting factors I meant, variables, and I have solved the problem by inverting them. It should be sorted. With regards to the factor loading, after some reading, I used a parallel analysis to determine I need two factors. The new problem now is that the FA model does not satisfy RSMEA 0.9. I only achieve the satisfaction of those two requirements when I use 4 factors. What would be the best to listen to, the parallel analysis or adding/removing factors until the FA model makes sense w.r.t. RSMEA & Tucker's. Thanks again for your input, very much appreciated.
@@TheBlomb. Haha yeah. Regarding which "better" method to use, I'd defer to the testing data. If there is a better "score" of some sorts, I'd more or less default to that specific method. However, if you want to avoid the black box algorithm output, and assuming they both have similar results, I'd go with the method that has the most explanatory power. Could you do some additional analysis and try and determine what the factors might represent your original features and determine if the representations make sense?
Hi, I'm having trouble using the data set in the link. It downloads as a .arff file and I've tried using both foreign and Weka packages to import the data with no success
Why would you consider absolute BIC values only? Whether the values are positive or negative, you should consider the model that gives you smaller BIC: see en.wikipedia.org/wiki/Bayesian_information_criterion. Please let me know if I misunderstood anything and thanks in advance for your input. And thanks for the great tutorial.
Great question! Sometimes, you have negative BIC values. You'd want to take the absolute value of these values to have an 'objective' valuation of this model with other models that you have. If you have a -inf BIC value.. this is as 'bad' as a +inf value.
I have a question. So, you arbitrarily stated the number of factors you used in this video and you said that a more 'scientific' way of determining numbers of factors could be done by using scree plot (elbow method). You know that some data don't have a good scree plot diagram, right? If that's the case, can I use like the Eigenvalue thing? Thank you so much!
Yes! You can use just the eigenvalues to determine if you want to keep X features. If the eigenvalue < 1 can be eliminated. However, this approach is more appropriate for smaller models and can restrict larger models.
while applying the pa and ml method the values for the h2 and u2 is not showing and also i have 10 variables but for 3 factors the cumulative variance is 0.43. so what to do?
What is the variance explained with you add up the 4th factor? Also, you might have NaN values in your dataset which may have led to not having any h2 and u2 values? (speculating)
@@gaoelnlaojehc8913 I appreciate your kind words. :) I usually start off at .edu websites. Stanford has a lot of great theory related to ML. Medium articles have application oriented scripts that can hammer home understanding. And, of course there is stack overflow.
Excelent video! Keep up with the good work, you are helping a lot of distressed students!
Very nicely explained
Good work!
목소리가 너무 매력적이에요! 우와
Ah thank you so much sir 🙏
Thank you, that was very helpful! ^^
thanks so much...I'm working on my thesis and this was extremely helpful... 🙃
how to use output of factor analysis output/factors in clustering or other ML algo like classification?
Great question!
Until I do a topic on Factor Analysis Regression (FCR), you can take a look at this video: th-cam.com/video/H45NWCzIDkY/w-d-xo.html
(I go over the logic on how to use the transforms for other modeling procedures -- its the same process)
After factor analysis, how we can create an index or indicator?
Thank you for this video! Do you have any suggestions on how to handle missing data in factor analysis?
I have just the video for you :) (data imputations)
th-cam.com/video/MpnxwNXGV-E/w-d-xo.html
How do you know the # of factors that you look for? Is this based from your research context, or can you read the is from the date by experimenting with more or less factors? & what if you have contrasting factor loadings, can you invert these? (I.E.: I have question on how challenging something is, and after how easy it is which is in contrast)
I arbitrarily stated that the number of factors be 3 in this video. However, a more "scientific" way of determining number of factors is either 1) using a scree plot (elbow method) or you can check out this paper here: ruscio.pages.tcnj.edu/files/2016/08/Ruscio-Roche-2012-PA-Factor-Analysis.pdf on a more mathy way of finding out number of factors to optimize for. (factor loadings)
You can also attempt to gridsearch for the number of factors (brute force based on model results)
I'm not sure what you mean by contrasting loading factors?
@@SpencerPaoHere Thank you for your response and the source. By contrasting factors I meant, variables, and I have solved the problem by inverting them. It should be sorted.
With regards to the factor loading, after some reading, I used a parallel analysis to determine I need two factors. The new problem now is that the FA model does not satisfy RSMEA 0.9. I only achieve the satisfaction of those two requirements when I use 4 factors. What would be the best to listen to, the parallel analysis or adding/removing factors until the FA model makes sense w.r.t. RSMEA & Tucker's.
Thanks again for your input, very much appreciated.
@@TheBlomb. Haha yeah. Regarding which "better" method to use, I'd defer to the testing data. If there is a better "score" of some sorts, I'd more or less default to that specific method.
However, if you want to avoid the black box algorithm output, and assuming they both have similar results, I'd go with the method that has the most explanatory power.
Could you do some additional analysis and try and determine what the factors might represent your original features and determine if the representations make sense?
Hi, I'm having trouble using the data set in the link. It downloads as a .arff file and I've tried using both foreign and Weka packages to import the data with no success
Fixed this. Have to manually change the data types in the .arff file to Numeric from integer
Do you have a video explaining Tucker’s? Thanks
Hmm I do not. I am not even sure what Tucker's is haha. Sounds very niche.
Why would you consider absolute BIC values only? Whether the values are positive or negative, you should consider the model that gives you smaller BIC: see en.wikipedia.org/wiki/Bayesian_information_criterion. Please let me know if I misunderstood anything and thanks in advance for your input. And thanks for the great tutorial.
Great question! Sometimes, you have negative BIC values. You'd want to take the absolute value of these values to have an 'objective' valuation of this model with other models that you have. If you have a -inf BIC value.. this is as 'bad' as a +inf value.
I have a question. So, you arbitrarily stated the number of factors you used in this video and you said that a more 'scientific' way of determining numbers of factors could be done by using scree plot (elbow method). You know that some data don't have a good scree plot diagram, right? If that's the case, can I use like the Eigenvalue thing? Thank you so much!
Yes! You can use just the eigenvalues to determine if you want to keep X features. If the eigenvalue < 1 can be eliminated. However, this approach is more appropriate for smaller models and can restrict larger models.
@@SpencerPaoHere Ah now I get it. THANK YOU SO MUCH! My lecturer had given me this material but I didn't get a single thing she said T_T
while applying the pa and ml method the values for the h2 and u2 is not showing and also i have 10 variables but for 3 factors the cumulative variance is 0.43. so what to do?
What is the variance explained with you add up the 4th factor? Also, you might have NaN values in your dataset which may have led to not having any h2 and u2 values? (speculating)
My analysis for
factor analysis is returning NA- not available
How can I help?
His voice
where did u learn bro?
The internet :P
@@SpencerPaoHere Any specific sites or blogs sir? BTW I like ur vids man... pls continue
@@gaoelnlaojehc8913 I appreciate your kind words. :) I usually start off at .edu websites. Stanford has a lot of great theory related to ML. Medium articles have application oriented scripts that can hammer home understanding. And, of course there is stack overflow.
@@SpencerPaoHere Thanks man. Appreciate it
Half of me is thinking "You're just making this all up" haha