Machine Learning for Information Management: Some Promising Directions

Management of personal information such as email messages, calendar entries, to-do items, and workstation documents is one of the most highly visible current uses of computer technology. I will present experimental evidence that machine learning techniques can be effectively used to improve personal information management tools in two ways. First, machine learning can be used to improve performance on certain types of difficult searches, notably searches that require some awareness of context. Second, machine learning can be used to reduce the chance of certain high-cost errors. One type of high-cost error we consider is the “dropped ball”—i.e., losing track of a task that has been delegated, in part or whole, to others. The second type of high-cost error is an “email leak”—i.e., mistakenly sending a sensitive email message to the wrong recipient.