Abstract
This paper describes an algorithm to cluster and segment sequences of low-level user actions into sequences of distinct high-level user tasks. The algorithm uses text contained in interface windows as evidence of the state of user-computer interaction. Window text is summarized using latent semantic indexing (LSI). Hierarchical models are built using expectation-maximization to represent users as macro models. User actions for each task are modeled with a micro model based on a Gaussian mixture model to represent the LSI space. The algorithm's performance is demonstrated in a test of web-browsing behavior, which also demonstrates the value of the temporal constraint provided by the macro model.