On Studying the Effect of Data Quality on Classification Performances

International audience ; The field of data repairing is very active and produces a lot of repairing methods. This abundance of options can make choosing a re- pairing method hard. Our research question stems from this problem: Is it always better to repair data? Our work is placed in the context of clas- sification tasks and numerical data. We investigates our research ques- tion through five criteria: C1 the perceived difficulty of using a repairing method according to experts, C2 the impact of the level of degradation of data on accuracies and f1 scores, C3 the effectiveness of the repair- ing tool, C4 the impact of the type of error present in data and C5 the impact of the classification model used. Our main contributions are a method to evaluate the difficulty of using a repairing tool and a study of experimental results on accuracies and f1 scores investigating C2 to C5. We were able to identify two categories of error types: low impact and high impact on accuracy and f1 score. We also observed that repair- ing methods perform similarly both at very low and very high levels of errors.