שמעון ליאור רוקח

Sequential data has been widely collected in the past few years; in the public health domain it appears as collections of medical events such as lab results, electronic chart records, or hospitalization transactions. Publicly available sequential datasets for research purposes promises new insights, such as understanding patient types, and recognizing emerging diseases. Unfortunately, the publication of sequential data presents a significant threat to users’ privacy. Since data owners prefer to avoid such risks, much of the collected data is currently unavailable to researchers. Existing anonymization techniques that aim at preserving sequential patterns lack two important features: handling long sequences and preserving occurrence times. In this paper, we address this challenge by employing an ensemble of Markovian models trained based on the source data. The ensemble takes several optional periodicity levels into consideration. Each model captures transitions between times and states according to shorter parts of the sequence, which is eventually reconstructed. Anonymity is provided by utilizing only elements of the model that guarantee differential privacy. Furthermore, we develop a solution for generating differentially private sequential data, which will bring us one step closer to publicly available medical datasets via sequential data. We applied this method to two real medical events datasets and received some encouraging results, demonstrating that the proposed method can be used to publish high quality anonymized data.