A Graph-Based Differentially Private Algorithm for Mining Frequent Sequential Patterns

Currently, individuals leave a digital trace of their activities when they use their smartphones, social media, mobile apps, credit card payments, Internet surfing profile, etc. These digital activities hide intrinsic usage patterns, which can be extracted using sequential pattern algorithms. Sequential pattern mining is a promising approach for discovering temporal regularities in huge and heterogeneous databases. These sequences represent individuals' common behavior and could contain sensitive information. Thus, sequential patterns should be sanitized to preserve individuals' privacy. Hence, many algorithms have been proposed to accomplish this task. However, these techniques add noise to the candidate support before they are validated as, frequently, and thus, they cannot be applied without having access to all the users' sequences data. In this paper, we propose a differential privacy graph-based technique for publishing frequent sequential patterns. It is applied at the post-processing stage; hence it may be used to protect frequent sequential patterns after they have been extracted, without the need to access all the users' sequences. To validate our proposal, we performed a detailed assessment of its utility as a pattern mining algorithm and calculated the impact of the sanitization mechanism on a recommender system. We further evaluated its information loss disclosure risk and performed a comparison with the DP-FSM algorithm.

Automated summary

This paper proposes a differential privacy graph-based technique for publishing frequent sequential patterns, which can protect these patterns without accessing all users' sequences. The utility of this technique as a pattern mining algorithm is assessed, along with its impact on a recommender system. A comparison with the DP-FSM algorithm is also performed.