Experiments in UNIX Command Prediction

Most users demonstrate regularities in their work with a computer system. Even when a user’s activities are unique, those interactions often exhibit systematic patterns. Accordingly, there has been a range of work developing systems that recognize regularities in computer usage (Cypher 1993; Mitchell et al. 1994; Schlimmer & Hermens 1993). Our current effort is to consider different methods that would enable us to build similar pattern-recognition into the UNIX command shell (Hirsh & Davison 1997; Davison & Hirsh 1997), although the concept of command prediction is similarly applicable to other user interfaces. The prediction of a user’s next command by the shell requires some mechanism that considers the user’s current session and from that information generates a prediction for the next command. We gather data by recording the sequence of commands executed as well as other pertinent information about the state of the shell. This forms the basis for applying an off-the-shelf inductive-learning method. To determine the potential for such methods, we collected command histories from over 70 users and computed online predictive accuracies for five methods on their data (Davison & Hirsh 1997). Online evaluation (train on the previous data points; test on the current one) was used instead of the traditional training and test set partitioning or crossvalidation because of the sequential nature of the data. We found that relatively straightforward, knowledge-free methods were able to correctly predict the next command (without arguments) that the user would execute up to 45% of the time. The best performance was achieved using C4.5 (Quinlan 1993) on two features (the two previous commands), and a training window of 1000 data points. Each subject contributed histories with, on average, 2184 commands, each with an average length of 3.77 characters. Assuming that a correct prediction could be inserted with a single character, this method would have saved just over 31% of the keystrokes typed, which is very close to the