Learning What a User Will Do Next:
A Case Study in UNIX Command-Line Prediction

Predicting what a computer-system user will do next would enable many useful applications and computer-system functionalities, such as more effective interfaces, speculative pre-execution of actions, intrusion detection, as well as potentially serving an important role in the design of more "personable" computer systems. In this talk I'll describe ongoing work on learning to predict a user's next action from the user's history of past actions with the computer system. In particular, we are exploring the level of success that fairly simple "knowledge-free" prediction methods can achieve, in contrast to more traditional "knowledge-based" methods. I'll describe the methods we are using, and our experiments using them on data from over 75 UNIX users gathered over a 2-6 month period.

(This is joint work with Brian Davison.)