A significant proportion of smartphone notifications are indicative of human behaviour (e.g. delivery updates for purchased items, physical activity summaries, and notification of updates to subscribed content). However, present attempts to understand human behaviour from smartphone traces typically focus on sensors such as location, accelerometer and proximity, overlooking the potential for notifications as a valuable data source. In this paper, we propose a general framework that provides end-to-end processing of notifications to understand behavioural aspects. We realise the framework with an implementation that tackles the specific use case of establishing prior buying behaviour from associated notifications. To evaluate the framework and implementation, we conduct a longitudinal user study in which we collect more than 250, 000 notifications, from twelve users, over an average of three months. We apply knowledge-based and machine learning techniques to those notifications to assess the tasks of the proposed framework. The results show a substantial difference in the performance between the methods used to extract behavioural features from the collected notifications.
|Journal||Behaviour \& Information Technology|
|Publication status||Published - 17 Nov 2022|
- digital phenotyping
- human behaviour
- machine learning