We study and review the techniques for de
aling with missing attribute values in data
mining. Then, we conduct the exper
iments to observe the performance
of classification algorithms on
each strategy of missing-value substitution. The
algorithms we used are naïve Bayes, tree-based
and instance-based classifiers. Four approaches
of handling missing values are introduced to the
numeric and nominal data sets
taken from the UCI repository. Th
e experimental results reveal the
superior suggestive choice of ignoring numerical
data instances with missing values, whereas
replacing the unknown values with the symbol “?”
produces better classification results for the
nominal data set.