RELATIVE MEMBERSHIP

In Table 3, the features generated are binary - i.e. we are checking for the presence of particular instantiated features. However, this gives a very ``crisp'' decision boundary. It may be more useful to have a richer measure of membership. We use the measure $D=log_2(\frac{d_2}{d_1})$, where $d_1$ is the distance to the nearest centroid and $d_2$ is the distance to the second nearest centroid. This measure has useful properties; for instance a point on the boundary between two regions has $D=0$, whereas the centroid has a measure of $D=\infty$. This expands the hypothesis language significantly and makes the classification more robust.