• E-step: “augment” data by predicting values of useful hidden variables
• M-step: exploit the “augmented data” to improve estimate of parameters (“improve” is guaranteed in terms of likelihood)

Properties:

• General algorithm for computing ML estimate of mixture models
• Hill-climbing, so can only converge to a local maximum (depending on initial points)

Examples:

• k-means, GMM
• pLSA
• LDA by variational inference
• ...

reference:
Text Mining: https://www.coursera.org/learn/text-mining