What Belongs Together Comes Together: Activity-centric Document Clustering for Information Work

Multitasking and interruptions in information work make frequent activity switches necessary. Individuals need to recall and restore earlier states of work which generally involves retrieval of information objects. To avoid resulting tooling time an activity-centric organization of information objects has been proposed. For each activity a collection with related information objects (like documents, websites etc.) is created to improve information access and serve as a memory aid. While the manual maintenance of such information collections is a tedious task and becomes an interruption on its own, the automatic maintenance of such collections using activity mining is promising. Activity mining utilizes interaction histories to extract unique activities based on the stream of interaction with information objects. For activity mining, existing work shows varying success in limited study setups. In this paper, we present a method for activity mining to generate activity-centric information object collections automatically from interaction histories. The technique is a hybrid approach considering all information types used in previous work -- activity stream and accessed content related information. Method performance is evaluated based on interaction histories collected during real work data from eight information workers collected over several weeks. For the dataset our hybrid approach shows on average a performance of 0.53 ARI up to 0.77 ARI, outperforming single metric-based approaches.