Conclusions

In conclusion, ID3 seemed the best decision tree method, considering the accuracy of predictions, consistency of predictions from the point when the learner gathered enough data to generate useful rules, and the small number of rules created. Perhaps the C4.5 methods’ ability to learn meaningful rules more quickly could be incorporated into a future learning assistant. It could use a C4.5 method as the personal assistant user is just beginning schedule events, then once sufficient amounts of meaningful data had been gathered or C4.5 learners are generating too many rules, it could use the ID3 decision tree.

It appears that eight weeks of data in this simple and predictable schedule would be a better starting time than after one month’s worth of scheduling data. A future experiment might begin at that threshold to gather more meaningful statistics.

An obvious improvement would be making a personal assistant that remembers successful rules and also eliminates duplicates if these rules that are remembered are persistent from one learning session to the next. Secondly, the first match of a rule may not be the best match of a rule, this rule finding algorithm might be improved by finding the best fitting rule. However such an endeavor would lead to discussion on “best fit” of the rule. A simple greatest number of attributes fit would be the first method to test, but this still leaves other methods to explore and compare to the greatest number fit.