K-Means

K-Means is a simple algorithm, the pseudocode for which can be found in Figure 4.11. However, initial settings for K-means (e.g. initial cluster membership and number of clusters) are set using problem-specific information. The K-Means algorithm implementation is largely historical, and is of limited practical use in the current version of TClass, its role being largely replaced by the E-M algorithm.

In our implementation, K-Means accepts the following parameters:

• numClusters: The number of clusters to create. It also accepts the value ``auto''. Since K-Means has no mechanism for estimating the number of clusters, some alternative means must be used. Using ``auto'' works out the number of clusters as the average number of instantiated features of the type we are interested in per training stream. For instance, if we look at Table 4.2, we see that there are a total of 12 instantiated features from 6 training streams, and hence it would calculate that there are $\frac{12}{6} = 2$ clusters of data. Default value is ``auto''.

• initialdist: Initial distribution of points (i.e., which clusters the points end up in) can be accomplished in two ways: random, where each point is allocated to a random cluster, or ordered. In the ordered method, just as we ``guessed'' at the number of clusters using the average number of instances, so we can also allocate cluster membership based on the order of events. For instance, if the number of clusters is calculated automatically to be two, and there are two events, then if over all of the data we put the first instantiated feature detected into the first cluster, and the second into the second cluster, then that would be a reasonable initial distribution of instantiated features. Default value is ordered.

• closeness: When computing the distance of instances for the evaluation of membership measures, should the distance from the centroid to the instantiated feature being considered be used, or should the distance to the nearest point belonging to the cluster be used? Default value is distance from the centroid.