Training a classifier requires a supply of example problems and the correct classification (label) for each. In some practical situations examples are plentiful, but obtaining labels for them is costly. Several algorithms exist for learning classification when only a small number of examples are "labelled" at the outset and the remainder are "unlabelled." This thesis presents continued work on the Guelph Cluster Class algorithm developed by Dara, Stacey and Kremer. Specifically, it investigates how the algorithm performs on ten real-world data sets over a range of parameter settings, and whether cluster validity indices can guide the setting of the parameters. An examination of a simple clustering problem points to explanations for the algorithm's behaviour, and tests of a variant algorithm that capitalizes on these observations are presented. Finally, this thesis explores whether clustering information can guide the selection of examples which, if labelled, would be especially informative for classifier training.