Exploring privacy prediction models through the analysis of shapley values.

Human mobility data are crucial for understanding mobility patterns and developing analytical services. Unfortunately, this type of data is sensitive. The correct identification of privacy risks is therefore essential before deciding to release the data to the public. Of equal importance is explaining to the user to whom the data relates what risks he or she faces and why, so that he or she can take part in this decision-making process. Some recent work has proposed the use of long short-term memory (LSTM) neural networks to predict privacy risk on raw trajectories and the use of SHAP for risk explanation. In this thesis we research an improvement of this approach by using recent algorithms in time series classification (ROCKET and InceptionTime) for risk prediction. In this way, we were able to achieve a better time efficiency than LSTM while having comparable performance. We also improve the SHAP explaination in two ways: first by tackling the problem of feasible computation of shapley values for the prediction of privacy risk on mobility trajectories. Second via interactive analyses and visualization of the shapley values to help the end-users understand why they are associated with a certain privacy risk and what they might be at risk for