We present the results of our experiment aimed to comprehensively understand the combination of 1) how smartphone users interact with their notifications, 2) what notification content is considered important, 3) the complex relationship between the interaction choices and content importance, and lastly 4) establish an intelligent method to predict user's preference to seeing an incoming notification. We use a dataset of notifications received by 40 anonymous users in-the-wild, which consists of 1) qualitative user-labelled information about their preferences on notification's contents, 2) notification source, and 3) the context in which the notification was received. We assess the effectiveness of personalised prediction models generated using a combination of self-reported content importance and contextual information. We uncover four distinct user types, based on the number of daily notifications and interaction choices. We showcase how usage traits of these groups highlight the requirement for notification filtering approaches, e.g., when specific users habitually neglect to manually filter out unimportant notifications. Our machine learning-based predictor, based on both contextual sensing and notification contents can predict the user's preference for successfully acknowledging an incoming notification with 91.1% mean accuracy, crucial for time-critical user engagement and interventions.
ASJC Scopus subject areas