Linearity I, Olin College of Engineering, Spring 2018 I will touch on eigenvalues, eigenvectors, covariance, variance, covariance matrices, principal component analysis process and interpretation, and singular value decomposition.
Nicely done. It hit the right level for someone who understands the linear algebra behind Eigenvectors and Eigenvalues but still needed to make the leap of connecting a dot or two in the application of PCA to a problem. Again, thank you!
Believe it or not, I've been wondering a lot about the concept of covariance because every video seems to miss the reason behind the idea. But I think I kind of figured it out today before watching this video and I drew the same exact thing that is in the thumbnail. So I guess was thinking correctly : ))
Great video! Can anyone tell how she decided that PC1 is spine length and PC2 is Body mass? Should we guess (hypothesize) this in real world scenarios?
Same as usual, right? Find lambda using det(Sigma - lambda * I) = 0, so just take lambda away from the main diagonal of the Cov. Matrix, take the determinant of that and you'd be left with some polynomial of lambda which you then solve for, each solution being a unique eigenvalue.
I do understand that eigenvalues represent the factor by which the eigenvectors are scaled, but how do they signify “the importance of certain behaviors in a system”, what other information do eigenvalues tell us other than a scaling factor? Also, why do eigenvectors point towards the spread of data?
If you consider a raw matrix or just geometric examples eigenvalues are just a scaling factor indeed. And you cannot say much more. But here, we are talking with additional context: we know we are doing statistics and putting "data" into a covariance matrix, which means we can now add more interpretations. The eigen vector is not just some eigenvector of some matrix, it's the eigenvector of a *covariance matrix* in the context of statistics, we've put data into a matrix whose elements measure all the possible spread of data, which is why we can now say an eigenvector points towards the spread of data and its length (eigenvalue) relates to the importance of that spread.
Around the minute of 1.36, you said "we divide by n for covariance", but we divide by n-1, instead. Please, do check on that. Thanks for the video. Maybe, I sohuld say estimated covariance has the n-1 division.
Find the Covariance Matrix of these variables, like at 2:15, and find its eigen decomposition (find its two dominant eigenvectors). The matrix at 5:30 is the two dominant eigenvectors. Each column is an eigenvector.
No one explains why they use covariance matrix. Why not use actual data and find its igen vector/igen values. I have been watching hundreds of videos books. No one explains that. It just doesn't make sense to me to use covariance matrix. Covariance is very useless parameter. It doesn't tell you much at all.
No it does especially using PCA. But you are right, you need actual data. Say the data are 3D points of some 3d objects, if you use this technique (build a cov matrix using the 3D points and do the PCA of it) then you will find a vector aligned with overall direction of the shape: for instance you will find the main axis of a 3d cylinder. This is quite a useful information.