Generative Classifier And Discriminative Classifier

13 Apr 2020

Recently I was learning the naive bayes classifier and learnt there are two categories of classifiers to estimate the distribution $P(Y|X)$ based on the given data, Generative Classifier and Discriminative Classifier.

I am puzzled on the differences between these two classifiers at the beginning, since their goals are same and their equations are similar. Besides, I also found there are two estimation methods, namely MAP (Maximum A Posteriori) and MLE(Maximum Likelihood Estimation) show up quite often when using these two classifiers. What’s the relationships between them?

So, I will try to sum up some of my findings during my exploring the correct path to understand these two classifiers and two estimation methods in this post.

 

Overview

There is a dataset $D=\{(\mathbf{x_1},y_1),…,(\mathbf{x_n},y_n)\}$ drawn from distribution $P(X,Y)$, and our purpose is to estimate $P(Y|X)$ so that we can predicate the label $y$ from any features $\mathbf{x}$.

Both Generative Classifier and Discriminative Classifier are used to learn the conditional probability $P(Y|X)$. But their way to learn that is different. A Generative Classifier first learns the joint probability $P(X, Y)$ and uses the Bayes Theorem to calculate the conditional probability $P(Y|X)$ while A Discriminative Classifier learns the $P(Y|X)$ directly.

And MLE and MAP are the method used to learn the joint probability or conditional probability we mentioned in that two classifiers.

 

Generative Classifier

In Generative Classifier, the joint probability $P(X, Y)$ is calculated by \begin{equation} P(X, Y) = \dfrac{P(Y)P(X|Y)}{\sum_{Y’}{P(Y’)P(X|Y’)}} \end{equation}

So we need to learn $P(Y)$ and $P(X|Y)$ first, the followings are the steps

1. Assume some functional form for $P(Y)$ and $P(X|Y)$

   • Assume $P(Y) \approx P(Y | \theta) = f(Y, \theta)$. $\theta$ represents the parameters or hypothesis we assumed for the real distributoin $P(X,Y)$.
   • There is a debat for the explanation of $\theta$ here between frequentist and Bayesian inference. Bayesian approach believes that $\theta$ is a random variable which can have its own distribution $P(\theta)$, meanwhile frequentist approach thinks $\theta$ is a inner unknown variable which associated with the distribution $P(X,Y)$.
   • $f$ is a function assumed by us.
   • The same to $P(X|Y)$, assume $P(X|Y) \approx P(X|Y, \theta) = g(X, Y, \theta)$.
2. Estimate parameters $\theta$ of $P(Y | \theta)$ and $P(X|Y, \theta)$ based on the dataset $D$.
   • We can use either MLE or MAP to learn
   • If use MLE, we are trying to find the $\theta = argmax_{\theta}{P(D|\theta)}$
   • If use MAP, we are trying to find the $\theta = argmax_{\theta}{P(\theta | D)} = argmax_{\theta}{P(D|\theta)P(\theta)}$
   • If use MAP, we will be Bayesian since we believe there is a distribution $P(\theta)$ for $\theta$.
3. Use Bayes rule to calculate $P(X|Y)$

 

Discriminative Classifier

In Discriminative Classifier, the conditional probability $P(Y|X)$ is learnt directly, the followings are the steps

1. Assume some functional form for $P(Y|X)$
   • Assume $P(Y|X) \approx P((Y|X, \theta) = f(X, Y, \theta)$.
2. Estimate parameters $\theta$ of $P(Y|X, \theta)$ based on the dataset $D$
   • We can use either MLE or MAP to learn too, the detail method is same as the methods in Generative Classifier.

 

Reference

https://machinelearningmastery.com/logistic-regression-with-maximum-likelihood-estimation/ http://www.cs.cornell.edu/courses/cs4780/2018fa/lectures/lecturenote04.html https://medium.com/@mlengineer/generative-and-discriminative-models-af5637a66a3