0

Maximum likelihood estimation considers the likelihood of data given parameters. In a non-parametric model like k-means clustering, can MLE still be used?

I know a Gaussian mixture model can use MLE to make a clustering algorithm, but it's parametric so I'm not interested in this case.

curiousgeorge
  • 299
  • 1
  • 7
  • Don't know about k-means, but maximum likelihood can be used in nonparametric contexts: https://stats.stackexchange.com/questions/110687/is-maximum-likelihood-estimation-mle-a-parametric-approach/110694#110694 There is also https://stats.stackexchange.com/questions/123230/k-means-and-maximum-likelihood and https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6062903/ ... – kjetil b halvorsen Aug 16 '20 at 02:08

1 Answers1

1

The premise of maximum likelihood estimation is that there is a probability distribution parameterized by variables. The goal is to find the variables that optimize the metric, which is the joint probability.

If you have a non-parametric model with no distribution, the metric is not a likelihood. In the case of k-means, it is some distance metric between your centers and data. So the terminology here is you are not doing a MLE, but minimization of a loss function.

The computational techniques to find optimal values are similar (eg. gradient descent algorithms) but MLE strictly refers to maximizing the defined joint probability.

Neil Liu
  • 36
  • 2