Hierarchical task instance mining in interaction histories

Knowledge work at computer workplaces involves execution of multiple concurrent tasks with frequent task interruptions. The complexity of the resulting work processes makes task externalization a desired goal towards facilitating analysis and support of knowledge work, e.g. by extracting and disseminating best practices.

In this paper, we present a task mining method that identifies tasks based on interaction histories. The method generates instances of a semantic hierarchical task model which captures an abstraction of the work processes. A specific characteristic of the method is that it mines tasks based on a combination of semantic and temporal features, extracted from enriched interaction histories. The use of semantic similarity results in a high robustness of the system with respect to task interruption and concurrent task execution. An evaluation of our task mining method based on a study with users executing frequently interrupted tasks is presented. One element of the evaluation is the assessment of different algorithms for semantic similarity computing, namely Term Matching (TM), Vector Space Model (VSM) and Latent Dirichlet Allocation (LDA). For an approach using VSM a precision of 0.83, a recall of 0.76 and a F1-measure of 0.79 is reached.