Discussion
A very important characteristic of ChiMerge is robustness. While it will sometimes do slightly worse than other algorithms, the user can be fairly confident that ChiMerge will seldom miss important intervals or choose an interval boundary when there is obviously a better choice. In contrast, the equal-width-intervals and equal-frequency-intervals methods can produce extremely poor discretizations for certain attributes, as discussed earlier. Another feature is ease of use; while
discretization quality is affected by parameter settings, choosing a x2-threshold between the .90 and .99 significance level and max-intervals to a moderate value (e.g., 5 to 15) will generally produce a good discretization (some qualifications are discussed later). A major
source of robustness is that unlike the simple methods,
ChiMerge takes the class of the examples into consider-
ation when constructing intervals and adjusts the num-
ber of intervals created according to the characteristics
of the data. In addition, ChiMerge is applicable to
multi-class learning (i.e., domains with more than two