Guassian mixure model. Exceeds k-means by the ability of handling oval-shaped clusters.

randomly initialize k centroids
loop
  assign cluster label with maximun likelyhood
  update mean and covariance matrix
until convergence

likelyhood (on the assumption that covariance matrix is diagonal):

$$\begin{align} p(x|c=k) &{}\propto N(\mu_k,\Sigma_k)\\ &{}\propto-\|x-\mu_k\|^2 \end{align}$$

Reference

"Machine Learning" Lecture 16: http://www.umiacs.umd.edu/~jbg/teaching/CSCI_5622/