4

LDA: linear discriminant analysis

Suppose we have a classification problem. I understand that the data can be such that the features may have discrete values or continuous values.

Suppose our data contains continuous feature values. Then we can apply Naive Bayes using a distribution. Lets assume the data to be normally distributed and so use Naive Bayes with normal distribution. We can also apply LDA which also uses Normal distribution.

Using Naive Bayes we assume the features to be independent and by using LDA we assume the covariance to be same for all the classes.

How does these assumptions make these 2 models perform differently and which is a better model and in which conditions?

user1825567
  • 1,336
  • 1
  • 12
  • 22
  • 2
    Just to be clear, are you referring to [linear discriminant analysis](https://en.wikipedia.org/wiki/Linear_discriminant_analysis)? – Emre Jun 04 '17 at 21:07
  • Yes..I should have written it – user1825567 Jun 05 '17 at 13:01
  • As a general way of reasoning, the performance drops you get depend on the assumptions violations: the more the assumptions of the models are respected, the better they work – Nicola Bernini Jun 04 '17 at 13:29
  • Can you please elaborate as to which is a more general model of Gaussian classifier and how are they different? – user1825567 Jun 05 '17 at 13:03

1 Answers1

1

As far as I know, Gaussian LDA and Gaussian Naive Bayes both assume the features follow normal distributions. However, GNB places a stronger assumption on feature independence, while Gaussian LDA does not require features to be independent. See this.

ebrahimi
  • 1,277
  • 7
  • 20
  • 39
R captian
  • 11
  • 1