MIMiS: Minimally Intrusive Mining of Smartphone User Behaviors

TL;DR

This paper introduces MIMiS, a framework that clusters smartphone user behaviors using minimal intrusive data, balancing privacy concerns with the ability to predict mental health states.

Contribution

It proposes the concept of privacy surfaces and an unsupervised clustering method to analyze user behavior with varying data intrusiveness levels.

Findings

MIMiS effectively clusters users with similar mental health scores.

Privacy surfaces can reduce data intrusiveness while maintaining predictive accuracy.

Clusters reveal behavioral patterns aligned with academic deadlines.

Abstract

How intrusive does a life-saving user-monitoring application really need to be? While most previous research was focused on analyzing mental state of users from social media and smartphones, there is little effort towards protecting user privacy in these analyses. A challenge in analyzing user behaviors is that not only is the data multi-dimensional with a myriad of user activities but these activities occur at varying temporal rates. The overarching question of our work is: Given a set of sensitive user features, what is the minimum amount of information required to group users with similar behavior? Furthermore, does this user behavior correlate with their mental state? Towards answering those questions, our contributions are two fold: we introduce the concept of privacy surfaces that combine sensitive user data at different levels of intrusiveness. As our second contribution, we introduce MIMiS, an unsupervised privacy-aware framework that clusters users in a given privacy surface configuration to homogeneous groups with respect to their temporal signature. In addition, we explore the trade-off between intrusiveness and prediction accuracy. MIMiS employs multi-set decomposition in order to deal with incompatible temporal granularities in user activities. We extensively evaluate MIMiS on real data. Across a variety of privacy surfaces, MIMiS identified groups that are highly homogeneous with respect to self-reported mental health scores. Finally, we conduct an in-depth exploration of the discovered clusters, identifying groups whose behavior is consistent with academic deadlines.