Combining Machine Learning and Statistical Disclosure Control to Promote Open Data

Abstract

We proposed machine learning-oriented statistical disclosure control as a novel solution for New Zealand Transport Agency to release more person-related variables in its open crash data for privacy-preserving data mining. Instead of making arbitrary decisions in variable aggregation and using perturbation to guard against reidentification attacks at the cost of data distortion, we creatively drew upon feature engineering and dimensionality reduction techniques such as correspondence analysis to make evidence-based data manipulation without distortion. In addition, we built random forest classifiers along the way to directly monitor the impact of our data manipulation on users’ modelling, rather than relying on traditional utility metrics such as information loss and difference of eigenvalues that are less interpretable for users. The dataset produced using our method satisfied 10-anonymity based on 11 quasi-identifiers, with less than 3% suppression, compared with only 3-anonymity based on no more than 8 quasi-identifiers with far more than 3% suppression commonly reported in literature. Furthermore, our method enabled random forest classifier to achieve 0.996 for AUC and 0.895 for F-score in predicting crash severity.