PrivTS: Differentially Private Frequent Time-Constrained Sequential Pattern Mining

Abstract

In this paper, we address the problem of mining time-constrained sequential patterns under the differential privacy framework. The mining of time-constrained sequential patterns from the sequence dataset has been widely studied, in which the transition time between adjacent items should not be too large to form frequent sequential patterns. A wide spectrum of applications can greatly benefit from such patterns, such as movement behavior analysis, targeted advertising, and POI recommendation. Improper releasing and use of such patterns could jeopardize the individually’s privacy, which motivates us to apply differential privacy to mining such patterns. It is a challenging task due to the inherent sequentiality and high complexity. Towards this end, we propose a two-phase algorithm PrivTS, which consists of sample-based filtering and count refining modules. The former takes advantage of an improved sparse vector technique to retrieve a set of potentially frequent sequential patterns. Utilizing this information, the latter computes their noisy supports and detects the final frequent patterns. Extensive experiments conducted on real-world datasets demonstrate that our approach maintains high utility while providing privacy guarantees.