One standard way to evaluate the quality of such techniques is to look at the distribution of correct clusterings versus incorrect clusterings. This can be quantified by looking at the [precision, recall and quality of the clusterings][1]... but assumes that you have a "ground truth" (i.e. that you know the real cluster the data point belongs to). Precision: $p = \frac{TP}{TP + FP}$, the percentage of positives that are true positives (and not mis-classified negatives). Recall: $r = \frac{TP}{TP + FN}$, the percentage of true positives that were correctly identified. Quality (AKA accuracy) = $q = \frac{TP + TN}{TP + TN + FP + FN}$, the percentage of correctly classified items from all classified items. Here $TP$ is the number of true positives, $TN$ the number of true negatives, $FP$ the number of false positives, $FN$ is the number of false negatives. If you have many clusters and need to evaluate their interaction, then [a confusion matrix may be more useful.][2] --- If you want to go into this in even more detail, I'd suggest reading the paper: > Andrew P. Bradley, The use of the area under the ROC curve in the evaluation of machine learning algorithms, Pattern Recognition, Volume 30, Issue 7, July 1997, Pages 1145-1159, ISSN 0031-3203, http://dx.doi.org/10.1016/S0031-3203(96)00142-2. (http://www.sciencedirect.com/science/article/pii/S0031320396001422) Keywords: The ROC curve; The area under the ROC curve (AUC); Accuracy measures; Cross-validation; Wilcoxon statistic; Standard error [A pre-publication copy of which is available.][3] [1]: https://en.wikipedia.org/wiki/Precision_and_recall [2]: http://en.wikipedia.org/wiki/Confusion_matrix [3]: http://espace.library.uq.edu.au/eserv.php?pid=UQ:8925&dsID=pr-t.pdf