Attention Manager

Overview

The initial goal of the Attention Manager is to learn when it's appropriate to interrupt the user. The Attention Manager classifies and learns user activity patterns and predicts the best opportunities to interrupt minimizing efficiency loss due to cognitive context swapping while maintaining user task focus. The Attention Manager will be used to message delivery to more opportune times. The next goal is to improve user performance through better understanding of RADAR as a tool, using learning algorithms.

The model "learns in the wild" by classifying data from desktop sensors. Our initial hypothesis is that the user is interruptible at task boundaries. Tasks are composed of a sequence of operations. We employ a two-phase learning approach to identify the hierarchy of operations and tasks. First we extract features from desktop sensors (e.g. mouse movements, keyboard activity, window events), and classify them into operations (e.g. editing email, opening address book, writing text, browsing the web, copying and pasting text in Word, and searching for a text string on a web page). The second phase of learning uses the Hidden Markov Models (HMM) for matching the operation sequences to identify tasks and task boundaries. Example tasks our system should be able to identify include responding to a request to send a document as a file attachment, composing an email, editing a spreadsheet, or searching for a paper on the web.

In addition, we will use invoking of and data from other RADAR systems (e.g. space-time planner, webmaster, etc.) to assist in task definition, improving the accuracy of our classifier. We will use data from RADAR user tests to fine-tune and evaluate our unsupervised learning system.

We are providing an API that gives access to the data collected by our sensors, and the operations and tasks learned by our machine learning model. The API will provide notifications of task boundaries as users reach them. This will be used to gate introduction of incoming email until the user has reached a task boundary. In addition, it will provide statistics about tasks and operations, such as their quantity and duration.