We will now apply the classification tree method to a published
dataset of 20 basalts from the Pindos Basin in Greece (Saccani et
al., 2003). This exercise will serve as an illustration of the way
the tree method works in practice, and of its limitations. The
features relevant to the classification tree analysis are shown in
Table 9. As a first example, consider sample GR 47b.
It travels through the full classification tree of Figure
4 as follows: TiO2.135%
Sr
156ppm
TiO
1.285%
MgO
9.595%
MgO
5.775%
Rb
3.675ppm. The same sample travels through the HFS tree of Figure
5 along the following path: TiO
2.135%
TiO
1.046%
Nd
12.785ppm. Note
that for the last step, the second surrogate variable was used (Table
1), because no isotopic ratios were measured for
these samples. No rare earth elements were measured for sample GR
56c. Therefore, its path through the HFS tree stops at node 4, where a
``follow the majority'' decision must be made. Since the distribution
of training data in this node is IAB/MORB/OIB = 73/194/43 (Table
2), GR 56c is classified as a MORB, albeit not
with the greatest confidence.
There is agreement between the full tree and the HFS tree for only half of the samples. Samples GR 50d, 51a and 51b were classified as IAB by the full tree, and as MORB by the HFS tree. The distinction between IAB and MORB is the hardest one to make. IABs have a much greater compositional diversity than both MORBs and OIBs. This is also reflected in most discrimination diagrams (see for example Figure 6). Furthermore, it might be possible that Mg was lost during the greenschist metamorphism that affected the Pindos ophiolites (Saccani et al., 2003). This would have caused sample GR 50d to be sent left, rather than right at node 7 of Figure 4. Likewise, it is possible that Sr-loss caused samples GR 71a-e and 195a to be sent left, rather than right at node 3 of the full tree. Finally, sample GR 181b was classified as an IAB by the HFS tree (Figure 5). However, its terminal node (left branch of node 5) is not very pure: IAB/MORB/OIB = 43/20/0, once again illustrating the difficulty of distinction between MORB and IAB affinities, which is caused by the complicated petrogenesis of the latter.