Bayes Theorem and ROC


Conditional Probability

Is the idea of understanding the probability an event A occurs given an event B. We formally say the Conditional Probability of A given B is defined as:

\[P(A | B ) = \frac{P(A \cap B)}{P(B)}\]

Bayes Theorem
Photo Courtesy of [“Conditional Probability and Independent Events”]

Pictorially we consider the 3 intersected events in $A \cap B$ in the new sample space $B \subset \Omega = 25 $ as all the only possible outcomes when A occurs given B. We easily see that $\frac{P(A \cap B)}{P(B)} = \frac{3}{25}$.

Bayes Theorem

Since we know that the conditional probabilities are defined as:

\[P(A | B ) = \frac{P(A \cap B)}{P(B)}\]


\[P(B | A ) = \frac{P(A \cap B)}{P(A)}\]

we can plug the second equation into the first and arrive at Bayes Theorem:

\[P(A | B ) = \frac{ P(B | A ) \cdot P(A)}{P(B)}\]

There are alternative forms of Bayes, but first consider this image.

Intersection of A and B
Photo Courtesy of [“Intersection of sets A and B”]

We can alternatively define a set B as the addition of the $B \cap A$ (purple) and $B \cap A^c$ (orange) elements, or $B = (B \cap A) \cup (B \cap A^c)$. Basically a set B is the portion that intersects with B and the portion that does not intersect with B.

Then we can write Bayes Theorem as:

\[P(A | B ) = \frac{ P(B | A ) \cdot P(A)}{P(B \cap A) + P(B \cap A^c)}\]

Or more compactly if you replace the two terms in the denominator again with bayes theorem.

i.e. $P(B \cap A) = P(B \mid A) \cdot P(A)$ and $P(B \cap A^c) = P(B \mid A^c) \cdot P(A^c)$:

We arrive at the alternate form of Bayes Theorem.

\[P(A | B ) = \frac{ P(B | A ) \cdot P(A)}{P(B | A ) \cdot P(A) + P(A^c) \cdot P(B | A^c)}\]

Binary Classification

When we consider two indicator Random Variables $X \sim Bern(P)$ and $Y \sim Bern(P)$, $X$ pertaining to ground truth observations and $Y$ a model mapping new observations to a set of known memberships (binary memberships in our case) then we can perform the task of classification.

If we want to evaluate the performance of our model we first need to define a few terms.

If X is our ground truth and Y is our classifier, then all the possible outcomes of our Random Variables are as follows:

  1. $X = \text{False} \cap Y = \text{True}$ False Positive. The model inaccurately determined an observation was positively a member of the the class we are interested in.

  2. $X = \text{True} \cap Y = \text{False}$ False Negative. The model inaccurately determined an observation was not a member of the the class we are interested in.

  3. $X = \text{True} \cap Y = \text{True}$ True Positive. The model accurately determined a positive membership.

  4. $X = \text{False} \cap Y = \text{False}$ True Negative. The model accurately determined a negative membership.

Prevalence ($\varpi$) is the measure interested in quantifying the probability or percentage of observations belong to the the membership we are interested in.

\[\varpi = P(X = \text{True})\]

Sensitivity ($\eta$) (Recall / True Positive Rate) is the measure of correctly classifying an observation to be a member of the class we are interested in. In other words, the Probability of a True Positive occurring out of a positive result.

\[\eta = P(Y = \text{True} \mid X = \text{True})\] \[= \frac{P(Y = \text{True} \cap X = \text{True})}{P(X = \text{True})}\] \[\eta = \frac{P(Y = \text{True} \cap X = \text{True})}{\varpi}\]

Specificity ($\theta$) (Selectivity / True Negative Rate) is conversely the measure of correctly classifying an observation to not be a member of the class we are interested in. The Probability of a True Negative.

\[\theta = P(Y = \text{False} \mid X = \text{False})\] \[= \frac{P(Y = \text{False} \cap X = \text{False})}{P(X = \text{False})}\] \[\theta = \frac{P(Y = \text{False} \cap X = \text{False})}{1 - \varpi}\]

Fall-out (False Positive Rate) is the probability that a negative sample is accurately classified.

\[= P(Y = \text{Positive} \mid X = \text{False})\] \[= 1 - P(Y = \text{False} \mid X = \text{False})\] \[= 1 - \theta\]

Reciever Operating Characteristic Curve (ROC)

Is a metric to evaluate the accuracy of a classifier based on the the $\frac{\text{True-Positive Rate}}{\text{False-Positive Rate}}$ $\big( \frac{\eta}{1 - \theta} \big)$ as functions of a parameterized function.

ROC Curve
Photo Courtesy of [“What is an ROC Curve”]

The top left corner of the ROC space corresponds to a perfect classifier. As the curve travels toward the right corner along the green line, you will notice the classifier’s tendency to classify things as positive until it always classify a sample as positive. Every point on the green line corresponds to the accuracy of flipping a coin.