Research Motivation
Research in frustration detection in the Affective Computing group at the
Media Lab has been inspired by the current disability of most human-computer
interfaces to make inferences about the affective state of a user who is
interacting with a computer. Consider the very familiar scenario of a (possibly
novice) user who runs into a difficulty when using unfamiliar software,
or when a piece of hardware fails to work properly. If the computer could
make inferences regarding the affective state of the user, it might use
that information to enhance the interaction with the user through its interface.
The issue of sensing and decoding affective expressions is an important
part of this research, and in this project, we approach the issue of analysis
and recognition of affective signals collected during an experimental procedure
in which the user confronted a "frustrating" scenario while interacting
with a computer.
Experimental Design
Research in the psychophysiology community has proposed that affective states
are mapped onto physiological states that we may be able to measure. In
order to collect data carrying affective content, an
experimental situation was designed so that a user, engaged in a computer
task, would experience a frustrating incident while his or her physiological
signals were being collected and recorded for further processing.
Results
Using the biophysiological signals collected during the experiment, we
proceeded to develop a user-dependent system that learns typical
patterns of the physiological signals under the stimulus of
frustration. Modeling affective signals is a hard problem. Some of the
difficulties we encounter are the uncertainty associated with the
effectiveness of the applied stimulus, and the non-stationarity of the
data. The analysis of the data set was carried out in a probabilistic
framework by implementing machine learning algorithms for analysis of
time series. Using Hidden Markov models, we have obtained recognition
rates which, for subjects with enough experimental data, are better
than random for 21 out of 24 of the subjects. For a more detailed
explanation of the data analysis techniques, you may download a
Postscript version of the Master's Thesis, Tech Report 446, on the TR
webpage.. The thesis is entitled "Stochastic Modeling of Physiological Signals with
Hidden Markov Models: A Step Toward Frustration Detection in
Human-Computer Interfaces", by Raul Fernandez, 1997.
Please send email to: galt@media.mit.edu. This page was last updated October 15, 1997.