PRODUCING COMPREHENSIBLE DESCRIPTIONS

If the learner used at the backend produces descriptions of the form rgnC = yes, then these can be used to create comprehensible descriptions by substituting the synthetic feature in place of the attribute name. Hence Figure 3 can be easily transformed into Figure 5 by substituting the parameters of C.

Figure: Human readable form of rules in Figure 3.

However, Figure 5 is lacking in one regard: what does ``approximately'' mean? Does timestep 10 mean between 9 and 11 or between 5 and 15? We can give an approximation of the bounds on these values by drawing a bounding box in the original parameter space of all the instances belonging to region C. Looking at Figure 2, we see that all the points in region C lie within the bounding box $d=[3,4], t=[9,12]$. Hence the rule in Figure 5 can be rewritten more clearly as shown in Figure 6. Note that this is not the same concept that the classifier uses on unseen instances, but it is still useful as an approximation. A slight modification of this approach allows it to be used with relative membership.

Figure: Adding bounds to Figure 5.