Illustration

Figure shown below depicts the boundary ellipses of the distribution surfaces (the scattering ellipses) for the first and second class, as well as the projections on feature axes (x_i, x_j) (yellow line depicts the projections of the union of the features' informativeness onto feature axes, and red depicts projections regarding the informativeness redundancy, i.e., 'common' informativeness). With the assumption that we have the same variance in both classes, we will have the same length of projections on feature axes.

We want to make the conclusions about:

• the single feature (variate)-x_i, or x_j, and for this case we are observing the length of intersection of the projections of ellipses on the corresponding axis, or
• pair of variates, and here we are looking at the degree of intersection of the scattering ellipses concluding that the amount of information we get is inversely proportional to the area of intersection, or the informativeness of joint features decreases with the increasing intersection of their distributions- which is not surprising, since we would find ourselves in great difficulty trying to separate classes with significant overlapping in feature space, and because of that

the main thing to do is to find the synergestic features, and everything is going to get crystal clear now, when you see the little animation (yeah, yeah, ... next page)

<< >>