An on-device gender prediction method for mobile users using representative wordsets

With the proliferation of mobile devices and the growing necessity for gender information in personalized intelligent systems, gender prediction of mobile users has become an important research issue. Text data in mobile devices are known to have high discriminative power for gender, but transmitting those data to the outside of a device has a security risk and raises a privacy concern of users. This study introduces an on-device gender prediction framework, by which the entire data analysis is performed inside a device minimizing the privacy risk. To cope with the resource limitation of mobile devices, gender information of a user is predicted by matching the user's mobile text data against gender representative wordsets which are constructed from web documents using a word evaluation measure. From the experiments conducted on real-world datasets, the effectiveness of the proposed framework was confirmed, and it was concluded that not only discriminability of a word but also popularity should be considered for the on-device gender prediction. The proposed framework is simple yet very powerful for gender prediction that its practical application to various expert and intelligent systems is possible attributed to the low computational complexity and high prediction performances.