The current trend in debugging and testing is to cross-check information collected during several executions. Jones et al., for example, propose to use the instruction coverage of passing and failing runs in order to visualize suspicious statements. This seems promising but lacks a formal justification. In this paper, we show that the method of Jones et al. can be re-interpreted as a data mining procedure. More particularly, the suspicion indicator they define can be rephrased in terms of well-known metrics of the data-mining domain. These metrics characterize association rules between data. With this formal framework we are able to explain limitations of the above indicator. Three significant hypotheses were implicit in the original work. Namely, 1) there exists at least one statement that can be considered as faulty ; 2) the values of the suspicion indicator for different statements should be independent from each others; 3) executing a faulty statement leads most of the time to a failure. We show that these hypotheses are hard to fulfill and that the link between the indicator and the correctness of a statement is not straightforward. The underlying idea of association rules is, nevertheless, still promising, and our conclusion emphasizes some possible tracks for improvement.
