This well-written paper reports on a study revealing the power of limited datasets to characterize mobile device users. This excellent work deserves attention both for the results and for the general explanations of technologies and use of data, as well as for its explanations of specific methodologies for working with subjects, statistics, and machine learning. The study was included in the PhD research of Kamini Garg, one of the authors.

The study recruited subjects to allow access to their mobile traffic for a limited period of time. The access included four profiling scenarios: (1) timestamps, (2) hypertext transfer protocol (HTTP) headers, (3) category of website as determined by the domain name, and (4) web page content. Note that no user input was included, but the HTTP header information does indicate when user interaction occurred.

The subjects completed surveys indicating demographics, personality traits, boredom proneness, and shopping interests (for example, computers and electronics, furniture, travel). The International Personality Item Pool (IPIP) questionnaire was used to capture the “big five” personality traits (dimensions of extroversion, neuroticism, agreeableness, conscientiousness, and openness).

The authors then used machine learning to determine how each of the datasets puts the users in high or low categories for target variables (multiple personality attributes, boredom proneness, demographics, and shopping). The profiling scenarios, most surprisingly the first two, performed well. The authors note that, given the power of what could be construed as non-personal information, it is important to inform users how data is being analyzed and for what purposes.