Performance comparison of the EM MST<i>k</i>NN with <i>k</i>-Means, SOM, CLICK and the original MST<i>k</i>NN algorithms, in terms of homogeneity and separation.

<p>The implementations of the <i>k</i>-Means, SOM, CLICK algorithms are obtained from the Expander microarray data cluster tool in <a href="http://www.plosone.org/article/info:doi/10.1371/journal.pone.0045535#pone.0045535-Sharan1" target="_blank">[124]</a>. The homogeneity and separation are computed using the definition in <a href="http://www.plosone.org/article/info:doi/10.1371/journal.pone.0045535#pone.0045535-Sharan1" target="_blank">[124]</a>. The AD ratio metafeatures data set is generated by taking pair-wise ratios between the features in 1,372-probe AD signatures <a href="http://www.plosone.org/article/info:doi/10.1371/journal.pone.0045535#pone.0045535-GomezRavetti1" target="_blank">[5]</a> and including MMSE score, NFT count, Braak staging, <i>JSD<sub>control</sub></i> and <i>JSD<sub>severe</sub></i> as five progression markers. The other data set contains four different types of metafeatures (ratios, summations, differences and products) and the aforementioned progression markers.</p>