# Maximum Likelihood Classifier vs. Bayesian Network

I am confused in understanding the Maximum Likelihood as a classifier. I know what is Bayesian network and I know that ML is used for estimating the parameters of models. Also, I read that there are two methods to learn the parameters of a Bayesian network: MLE and Bayesian estimator. The question which confused me is that:

1. Can we use ML as a classifier? For example, can we use ML to model the user' behaviors to identify the activity of them? If yes, How? What is the likelihood function that should be optimized? Should I suppose a normal distribution of users and optimize it?

2. If ML can be used as a classifier, what is the difference between ML and BN to classify activities? what is the advantages and disadvantages of each model?

• Hi Atena and welcome to this community. Please, ask just one question per post! – nbro Sep 13 at 22:09

## 1 Answer

If you read nothing else, maximum likelihood estimate => chance that the data predicted is the data observed. If you have a range of points (2, 3, 4, 5, 71) your MLE is going to favour ~4.5 because of means and standard deviations. MLE speeds up finding good input parameters, usually for a different classifier.

To answer your question:

1) Columbia University have a great example of using MLE classifiers, where everything is broken down into bitesize (or bytesize) chunks. Read this. Seriously.

2) In short, MLE is best used for simple, univariable distributions. It doesn't scale well to big problems but is waaay faster than a Bayesian network for simple tasks like predicting your height based on the heights of your immediate relatives. If you want to get technical, the conditional probability network of the Bayesian model reveals insights faster than the chain multiplication of the more primitive MLE.

Hope it helps!