Section: New Results
Inference of human personality from mobile phones datasets
Participants : Adriano Di Luzio, Aline Carneiro Viana, Julinda Stefa, Katia Jaffres-Runser [INPT-ENSEEIHT - IRIT (Toulouse University)] , Alessandro Mei [Sapienza University (Italy) - Dept. of Computer Science] .
Related to human behavioral studies, personality prediction research has enjoyed a strong resurgence over the past decade. Due to the recognition that personality is predictive of a wide range of behavioral and social outcomes, the human migration to the digital environment renders also possible to base prediction of individual personality traits on digital records (i.e., datasets) mirroring human behaviors. In psychology, one of the most commonly used personality model is the Big5, based on five crucial traits and commonly abbreviated as OCEAN: Openness (O), Conscientiousness (C), Extroversion (E), Agreeableness (A), and Neuroticism (N). They are relatively stable over time, differ across individuals, and, most importantly, guide our emotions and our reactions to life circumstances. It is so for social and work situations, and even for things as simple as the way we use our smartphone. For instance, a person that is curious and open to new experiences will tend to look continuously for new places to visit and thrills to experience.
This work brings the deepest investigation in the literature on the prediction of human personality (i.e., captured by the Big5 traits) from smartphone data describing daily routines and habits of individuals. We take a ground-breaking step in (i) deeply capturing human habits in terms of movements, visits, wireless connectivity as well as some routinary actions from a crowdsourced mobility dataset and in (ii) better understanding the relationship between personality traits and individual behavior. We do so by leveraging a dataset collecting very detailed routines of individuals originating from different countries located in 2 different continents, who answered the Big Five Inventory and allowed continuous collection of data from their smartphones for research purposes for 3 years. We use this dataset to engineer a set of human-adapted features that capture three aspects of human behavior: Temporal Mobility (e.g. time at home/work or commuting), Spatial Mobility (e.g. number of most frequent places, maximum distance from home), and the Context of Use (battery charging habits, wireless hotspots availabilities). Then, we use the features that have a statistically significant correlation with the OCEAN traits to predict the personality of a test-set portion of our dataset through cross validation.
Our results attest an accurate prediction of users' personality traits when a 5-level granularity is used per trait. This brings a much higher precision to our predicted results, when compared to the usual 3-level literature granularity. In addition, our prediction methodology carefully takes advantage of engineered features that (1) are more human-adapted and consequently, allow better capturing individuals' habits in terms of movements, visits, connectivity, context, as well as actions (note that contrarily to the literature, neither calls behavior nor data content is leveraged in our analysis), and (2) are designed having in mind the differences and particularities among the Big5 traits of personality. Thus, this work has the potential to impact the way we characterise unique behaviors of individuals as well as quantify how human personality influences lives and actions. Our results show (1) a significant correlation of most of the traits with a small set of mobility-related features and (2) that we are able to predict the individuals' Big5 traits with considerable accuracy (e.g., prediction of the 5 levels of Openness trait shows an F1 score of 0.77), which is significantly outperforming a benchmark approach, when only considering a set of only 3 of our human-adapted features. Finally, we discuss the ethical concerns of our work, its privacy implications, and ways to tradeoff privacy and benefits.
This is an on-going work with Adriano di Luzio, who spent 4 months in our team working as an internship, Julinda Stefa, an invited research visitor at Infine, and two other researchers: Katia Jaffres-Runser and Alessandro Mei. A paper describing this work is under submission at ACM Mobihoc 2018, but a technical report is also registered under the name hal-01954733.