Summary

A small percentage of the rules generated were actually used. We can refer to persistent rules as those that the learners re-generate on subsequent learning processes. Since the personal assistant forgets the previous rules, the new rules are the rules used to predict event preferences. Persistent rules made up a fair percentage of the rules regenerated from week to week. This was expected since each new learner only adds a few instances for the week of scheduled events when it relearns the rules. A small percentage of these persistent rules account for weeks 9, 10 and 11’s successes at predictions.

At most 4 rules were used to predict an attribute value. Most often, no rules were used, especially in the C4.5 cases. It is easily seen that the C4.5 methods both quickly created many rules, over 100. This slows down the personal assistant as all rules are checked until one is matched. Since only at most 4 and on average 2-3 rules, for those weeks when predictions were made, the overhead of rule creation and rule checking is wasting resources for this implementation.

ID3 prediction percents were all between 50 and 100 percent, where C4.5 for both pruned decision trees and unpruned decision trees, which ranged from 25 to 100 percent accurate predictions.

The ID3 method proved the most consistent learner of the three. ID3 appears to be better with smaller amounts of data, even from week to week, adding only a few instances improved the predictions of all the attributes. For the time and duration attributes the C4.5 methods were able to make predictions earlier, but the C4.5 pruned decision tree learner was not as consistent its unpruned counterpart. For the time ID3 was more consistent even though more time was needed to commence predictions. For duration, C4.5 unpruned is clearly the best method, but it is difficulty to say which of ID3 or C4.5 pruned are next in performance.