We began by considering how the number of divisions affected the results. The results are shown in figure 5.7.
Figure 5.7: The variation in the error rate that occurs
with segmentation.
This shows some recurring patterns. Again, IBL1 is very significantly better than C4.5.
We notice that the minimum occurs around 5 or 6 segments. We will use five, simply because it reduces the feature count by 8 features, and appears to give similar error rates.
Table 5.7: Results of learning algorithms on simple time division features.