Privacy and data mining: evaluating the impact of data anonymization on classification algorithms

Data anonymization is a technique used to increase the assurance that private data is not accessible to third parties. In data mining processes, anonymization can impact the results, since anonymized data may hinder the data analysis performed by algorithms commonly used in this context. The goal of this Practical Experience Report is to evaluate the accuracy and performance impact of data anonymization on data mining classifiers results. This is done through comparisons of their execution using original and anonymized data. A sample of real data generated by a Brazilian city transportation system associated to fictitious users was anonymized at different stages and classification algorithms, such as ZeroR, KNN (k-Nearest Neighbor), and Naive Bayes, were applied. Contrary to expectations, when the anonymization techniques were introduced in some classes, the accuracy was raised, as well as performance, reducing execution time. These results demonstrate that data anonymization techniques, when properly applied, can contribute to the effectiveness of data mining classifiers.