Predicting Personality Traits Using Smartphone Sensor Data and App Usage Data

Human behavior is complex -- often defying explanation using traditional mathematical models. To simplify modeling, researchers often create intermediate psychological models to capture aspects of human behavior. These intermediate forms, such as those gleaned from personality inventories, are typically validated using standard survey instruments, and often correlate with behavior. Typically these constructs are used to predict stylized aspects of behavior. Novel sensing systems have made tracking behavior possible with unprecedented fidelity, posing the question as to whether the inverse process is possible: that is, inferring psychological constructs for individuals from behavioral data. Modern smartphones contain an array of sensors which can be filtered, combined, and analyzed to provide abstract measures of human behavior. Being able to extract a personal profile or personality type from data directly obtainable from a mobile phone without participant interaction could have applications for marketing or for initiating social or health interventions. In this work, we attempt to model a particularly salient and well-established personality inventory, the Big Five framework. Daily routines of participants were measured from parameters readily available from smartphones and supervised machine learning was used to create a model from that data. Cross validation-based evaluation demonstrated that the root mean squared error was sufficiently small to make actionable predictions about a person's personality from smartphone logs, but the model performed poorly for personality outliers