A Comparison of Mining Incomplete and Inconsistent Data

We present experimental results on a comparison of incompleteness and inconsistency. Our experiments were conducted on 141 data sets, including 71 incomplete data and 62 inconsistent, created from eight original numerical data sets. We used the Modified Learning from Examples Module version 2 (MLEM2) rule induction algorithm for data mining. Among eight types of data sets combined with three kinds of probabilistic approximations used in experiments, in 12 out of 24 combinations the error rate, computed as a result of ten-fold cross validation, was smaller for inconsistent data (two-tailed test, 5 % significance level). For one data set, combined with all three probabilistic approximations, the error rate was smaller for incomplete data. For remaining nine combinations the difference in performance was statistically insignificant. Thus, we may claim that there is some experimental evidence that incompleteness is generally worse than inconsistency for data mining.