Modeling the Temporal Nature of Human Behavior for Demographics Prediction

Mobile phone metadata are increasingly used to study human behavior at largescale. In developing countries, where data is scarce, data generated by mobile phones have already helped with planning disaster response and informing public health policy. Basic demographic information such as age or gender is, however, often missing from mobile phone data. There has recently been a growing interest in predicting demographic information from metadata. Previous approaches relied on standard machine learning algorithms and hand-engineered features. We here apply, for the first time, deep learning methods to mobile phone metadata using a convolutional network. We represent the data as 8 matrices summarizing incoming and outgoing mobile phone usage (contacts, calls, texts, call duration) on a given week. Our method outperforms the previous state of the art on both age and gender prediction. Our analyzes indicate that our method captures compositional hierarchies in the mobile metadata based on the temporal nature of the data. These results show great potential for deep learning approaches for prediction tasks using standard mobile phone metadata.