I have read many posts on the topic (e.g. this or this) and have gone through a few introductions/tutorials, however, while I do understand the mathematical description, I still have a lot of difficulty understanding the results of CCA on an intuitive level.

Basically, assuming the notation

$X \in \mathbb{R}^{n\times d}, Y \in \mathbb{R}^{n\times p}$

for two different data/features matrices, I understand that the goal of CCA is to find the transformations $\textbf{a} \in \mathbb{R}^d, \textbf{b} \in \mathbb{R}^p$ such that $X\textbf{a}$ and $Y\textbf{b}$ are maximally correlated. However, once $\textbf{a}$ and $\textbf{b}$ are found, I fail to understand what the information that they reveal about the original feature set is (and therefore, I think that I'm missing the motivation of CCA). For example, consider a toy example such as the one illustrated by the figure,

fig. 1

where, keeping the definitions for $X$, $Y$ introduced earlier and using Matlab-style notation,

# we generate 50 examples, each with 4 features that we devide into
# two groups, i.e. p=2 and q=2 in the above.


The last figure is the result of plotting the first canonical variables, and as expected, we see that the resulting variables have a correlation that is close to $1$.

All of this is great, but... Well... Now what? In particular:

  1. What does that tell about the original set of variables/features/descriptors (i.e. the columns of $X$ and $Y$)?

  2. What is the benefit of working with the basis that give us a diagonal correlation matrix?

  3. What if the first canonical variables had low correlation? What would be the interpretation of that?

PS: I mostly work in computer vision, so please excuse me for the inevitable misuse of jargon and approximations.



Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge that you have read and understand our privacy policy and code of conduct.

Browse other questions tagged or ask your own question.