Prediction and Discovery of Users' Desktop Behavior

We investigate prediction and discovery of user desktop activities. The techniques we explore are unsupervised. In the first part of the paper, we show that efficient many-class learning can perform well for action prediction in the Unix domain, significantly improving over previously published results. This finding is promising for various human-computer interaction scenarios where rich predictive features of different types may be available and where there can be substantial nonstationarity. In the second part, we briefly explore techniques for extracting salient activity patterns or motifs. Such motifs are useful in obtaining insights into user behavior, automated discovery of (often interleaved) highlevel tasks, and activity tracking and prediction.