Studying the factors influencing automatic user task detection on the computer desktop

Supporting learning activities during work has gained momentum for organizations since work-integrated learning (WIL) has been shown to increase productivity of knowledge workers. WIL aims at fostering learning at the workplace, during work, for enhancing task performance. A key challenge for enabling task-specific, contextualized, personalized learning and work support is to automatically detect the user's task. In this paper we utilize our ontology-based user task detection approach for studying the factors influencing task detection performance. We describe three laboratory experiments we have performed in two domains including over 40 users and more than 500 recorded task executions. The insights gained from our evaluation are: (i) the J48 decision tree and Naïve Bayes classifiers perform best, (ii) six features can be isolated, which provide good classification accuracy, (iii) knowledge-intensive tasks can be classified as well as routine tasks and (iv) a classifier trained by experts on standardized tasks can be used to classify users' personal tasks.