Improving Prediction of Real-Time Loneliness and Companionship Type Using Geosocial Features of Personal Smartphone Data

Improving Prediction of Real-Time Loneliness and Companionship Type Using Geosocial Features of Personal Smartphone Data. Congyu Wu et al. arXiv October 2020, https://arxiv.org/pdf/2010.09807.pdf

Abstract: Loneliness is a widely affecting mental health symptom and can be mediated by and co-vary with patterns of social exposure. Using momentary survey and smartphone sensing data collected from 129 Androidusing college student participants over three weeks, we (1) investigate and uncover the relations between momentary loneliness experience and companionship type and (2) propose and validate novel geosocial features of smartphone-based Bluetooth and GPS data for predicting loneliness and companionship type in real time. We base our features on intuitions characterizing the quantity and spatiotemporal predictability of an individual’s Bluetooth encounters and GPS location clusters to capture personal significance of social exposure scenarios conditional on their temporal distribution and geographic patterns. We examine our features’ statistical correlation with momentary loneliness through regression analyses and evaluate their predictive power using a sliding window prediction procedure. Our features achieved significant performance improvement compared to baseline for predicting both momentary loneliness and companionship type, with the effect stronger for the loneliness prediction task. As such we recommend incorporation and further evaluation of our geosocial features proposed in this study in future mental health sensing and context-aware computing applications.

8 Discussion

In this section we reflect on our outcome variables and approach in the grander context of understanding human behavior and enhancing human well-being through mobile sensing and data analytics. 

Temporal resolution The two related outcomes examined in this paper, loneliness and companionship type, fall in two overlapping yet distinguishable areas in ubiquitous computing research, namely mental health sensing and context-aware computing, respectively. Context-aware computing emphasizes a computer’s inference of its user’s activity and surroundings in real-time, thus naturally having a moment-to-moment granularity. However, mental health sensing tasks span a wider range of temporal resolutions. On the low end, we see condition diagnosis tasks observe participants for as long as two months consecutively and then offer a judgment about whether a participant is with a clinical condition such as depression. On the high end reside real-time tracking tasks like the one presented in this paper, which do not aim at a medical diagnosis but focus on raising timely warnings. In the middle of the scale, a number of studies have adopted temporal resolutions ranging from daily and every few days to weekly and bi-weekly. The differences in temporal resolution points to different types, formats, and content of intervention: following a diagnosis, traditional intervention programs may be applied as treatment, whereas predictions of higher temporal resolutions will enable just-in-time adaptive intervention via mobile platforms. Question as to what sensing-intervention scheme will be most efficacious for what cohorts and conditions remains open, challenging, and critical for successful future applications of smart mental health.

Social context Companionship type is a key aspect of an individual’s social context, but far from the entire picture. The extent to which companionship type was captured in this paper covers the existence of a companion and (if true) the nature of a companion but does not consider the number of people surrounding a participant, differences in distance, and the interaction behavior, which altogether constitute a holistic social context in which one is situated. To combat the arbitrariness in defining social context seen in extant literature and to systematically delineate the various aspects of social context sensing, we argue that a formalized response variable definition for future social context inference tasks is needed. We propose that four components, quantity, quality, distance, and interaction, be specified in a definition of social context in future context-aware computing work. Quantity refers to the number of individuals and quality refers to their social significance. The distance element, can be categorized into groups such as “within personal space”, “within social space”, and “beyond social space” based on Edward Hall’s proxemics theory [10]. The interaction element defines the type of in-person verbal interaction taking place, 18 which may include absence of interaction, interaction among others only, interaction involving self. Such a 4-pronged taxonomy will also help phrasing EMA questions to acquire ground truth in future sensing studies: as opposed to only asking “who are you with”, more detailed and rigorous questions may be administered.

Sensing hardware In this paper our core approach is feature engineering, utilizing Bluetooth and GPS data from Android smartphones. The capability of feature engineering in human-centric sensing and inference is inevitably bounded by both (a) the availability and degree of integration of a sensor and (b) the absolute content a sensor captures. In our large participant cohort, 88% were iPhone users, from whom Bluetooth data were unavailable; therefore to further utilize the predictive power of Bluetooth data in mental health sensing and context-aware computing practice, other wearable devices such as smart watches may provide a better habitat for relevant data processing and analytics. In existing literature on social behavior inference, Bluetooth data is the most utilized smartphone sensor but it is not nearly sufficient to distinguish finer grained scenarios such as the social contexts defined with the four components proposed in the previous paragraph. Introduction and fusion of novel or previously overlooked mobile sensors may offer new and more effective solutions to social context detection. Magnetometer and audio sensing are candidate options, as we are observing recent studies using phone-embedded magnetometer to detect coexistence for epidemiology applications [13] as well as ongoing work on wearable voice sensors [14], which have the potential to support emotional state prediction in daily life.