For our experiments, we used the Fall datasets from 1998, 1999 and 2000. The Spring dataset was not used as it would have required changes in the objective function to be accurate.
The 1998 dataset had not benefitted from DAO's extensive data integrity checking, unlike the later years. Essentially all the data entry errors were eliminated in the 2000 dataset, and less than a dozen problems with contradictory music tastes and birthdate errors remained in the 1999 dataset. There were 457 empty preferences for room type and first choice halls, and 23 first hall choice and special living option mismatches in 1998. Smoking students were the source of most of the contradictory preferences each year as they requested non-smoking halls, non-smoking special living options and non-smoking roommates:
|Year||Smokers||% NS Hall||% NS SLO||# NS Mates|