The generation of new Support Vectors (defining the clusters’
boundaries) allows us to define the cluster affiliation of
each feature vector. Therefore, we label the clusters and assign
each xti
2 Ft to its affiliated cluster using the Complete Graph
(CG) labeling method [16]. In CG a pair of feature vectors
xti
and xt
j is said to belong to the same cluster (connected
components) if the line segment (inside the feature space)
that connects them lies inside the hypersphere. A number
of feature vectors (usually 10 [16]) are sampled on the line
segment to assess the connectivity. We use 10 sampled feature
vectors as this number provides cluster labelling results with a
negligible error along with a limited computational complexity
[16]. A higher accuracy can be achieved with a larger number
of sampled feature vectors, but with the disadvantage of a
multiplicative increase in the computation time [44].
Therefore, we construct an adjacency matrix At =
[At
i;j ]NtNt where At
i;j = 1 if xti
and xt
j are connected
components. Each At
i;j is defined as