Finite mixture models have been explored for supervised, semi-supervised, and unsupervised classification. Fractionally supervised classification generalizes these approaches using the weighted likelihood. The weights in these models dictate the amount of supervision with 0, 0.5, and 1 corresponding to unsupervised, semi-supervised, and supervised classification, respectively. Weight values larger than 0.5 allow unlabelled data to inform the classifier more heavily than labelled data while values smaller than 0.5 do the opposite.

As demonstrated on benchmark clustering data sets, improvements in classification performance can be achieved by allowing weight values to deviate from these three special cases. Potential extensions and the complicated matter of weight specification will also be discussed.