6
$\begingroup$

In supervised learning, we refer to the regressors as independent variables and response variables as dependent, but from a probabilistic standpoint, I am having trouble understanding this.

To breakdown my confusion, I think it makes sense to consider two separate cases (1) regressors are fixed / constant / deterministic (2) Regressors are random variables


(1)

Constants can also be viewed as random variables. We know from probability theory that a constant random variable is independent of any other random variable and we also know that independence is symmetric. So if $X$ is independent of $Y$, then $Y$ is independent of $X$. You can see this easily from conditional probability $P(X,Y) = P(X|Y)P(Y) = P(Y|X)P(Y)$. So if $X$ is independent of $Y$, then we have $P(X|Y) = P(X)$. So $P(Y|X)$ must be $P(Y)$.

But how does this make sense in the context of supervised learning? We assume that $Y$ is dependent on $X$, but not vice versa?


(2)

The same idea holds as the above except $X$ is no longer fixed here.

$\endgroup$

1 Answer 1

4
$\begingroup$

The "dependent" and "independent" terminology for the variables is unfortunate terminology, which is best avoided. Statistical dependence is always bidirectional ---i.e., if a variable is statistically dependent on another variable, then that second variable is also statistically dependent with the first variable. In a regression model the two variables are posited to have a statistical relationship. We treat the explanatory (regressor) variables $\mathbf{x}$ as fixed and we model the regression function $u(\mathbf{x}) = \mathbb{E}(Y|\mathbf{x})$, which is the conditional expected value of the response (regressand) variable $Y$. See this related question for more discussion on the unfortunate terminology.

$\endgroup$

Your Answer

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

Not the answer you're looking for? Browse other questions tagged or ask your own question.