Information Technology Reference
In-Depth Information
Fig. 4.2 Hierarchical classification from ten ICA mixtures (Laplacian and uniform source
distributions). The bottom-up construction of higher-level clusters is indicated. a Data and basis
vectors for each ICA mixture at level h ¼ 1, b clusters at level h ¼ 5, c clusters at level h ¼ 9
Fig. 4.3 Dendrogram with
KL-distances between
clusters at each merging level
Fig. 4.4 and following the criterion of coefficients PC and PE explained above, the
optimum number of clusters was obtained at level h = 5 (i.e., grouping the data in 6
clusters as shown in Fig. 4.2 b).
The first step of the proposed hierarchical algorithm is to estimate the parameters
of the ICA mixtures that will determine the clusters at the lowest level of the
hierarchy. Thus, errors in the source densities of the model at this level will
propagate to higher levels producing erroneous groupings. In order to measure the
Search WWH ::




Custom Search