An empirical comparison between stochastic and deterministic centroid initialisation for K-Means variations

A semi-supervised sparse K-Means algorithm

“…We have also showed that, similar to unsupervised methods [15], semi-supervised algorithms can be affected by initialisation procedures and by the type of used constraints. From our results we observed that using only MUST-LINK constraints has a negative effect on the semi-supervised algorithms (however PCSKM could cope with the MUST-LINK constraints much better than MPCKM and PCKM).…”

“…Our benchmark includes the real world data sets fisheriris and ionosphere from the UCI repository [18] which have unknown feature quality. We also added two synthetic data sets that were generated based on the generator of [17] which we used in our previous study [15]. These synthetic data sets are consisting of informative and uninformative features without any noise injection.…”

“…In our first experiment we wanted to assess the performance of PCSKM compared with other unsupervised algorithms (LKM and SKM) and semi-supervised algorithms (PCKM and MPCKM). We tested the performance of all these algorithms using the deterministic initialisation technique of DKM++ [23] which on average performed best based in our previous benchmark [15], but does not take into account constraints. We also tested the semi-supervised algorithms with different number and types of constraints including, only MUST-LINK, only CANNOT-LINK, and random selection from both MUST-LINK and CANNOT-LINK.…”

“…We name this algorithm Pairwise Constrained Sparse K-Means (PCSKM) and we testing its performance under different conditions such as different number and kind of constraints (CANNOT-LINK, MUST-LINK or both). In our previous study [15] we have shown that the deterministic initialisation method of Density K-Means++ (DKM++) on average performs better than stochastic methods thus we select this method for the initialisation of the algorithms along with the seeding method proposed in the study of [16]. In our benchmark we include synthetic data sets from the study of [17] with known feature quality and real world data sets from the UCI [18] and a real world data set from our previous behavioural neuroscience study [19] which contains ten known uninformative features.…”

A semi-supervised sparse K-Means algorithm

“…We name this algorithm Pairwise Constrained Sparse K-Means (PCSKM) and we test its performance under different conditions such as different number and kind of constraints (CANNOT-LINK, MUST-LINK or both). In our previous study [14] we have shown that the deterministic initialisation method of Density K-Means++ (DKM++) [15] surpasses the average performance of stochastic methods thus we select this method for the initialisation of the algorithms along with the Seeding method proposed in the study of [16]. We have also included the initialisation methods of ROBIN [17] and Maximin [18] to strengthen our conclusions (results in appendix).…”