"Affective Understanding:" Modeling and Responding to User Affect

Once the Sensing and Recognition modules have made their best attempt to translate user signals into patterns that signify the user's emotional responses, the system may now be said to be primitively aware of the user's immediate emotional state. But what can be done with this information? How will applications be able to make sense of this moment-to-moment update on the user's emotional state, and make use of it? The Affective Understanding module will use, process, and store this information, to build and maintain a model of the user's emotional life in different levels of granularity--from quick, specific combinations of affective responses--to meta-patterns of moods and other emotional responses. This module will communicate knowledge from this model with the other modules in the system.

The Affective Understanding module will eventually be able to incorporate contextual information about the user and his/her environment, to generate appropriate responses to the user that incorporate the user's emotional state, the user's cognitive abilities, and his/her environmental situation.

Features of the module

The Affective Understanding module may:

• Absorb information, by receiving a constant data stream on the user's current emotional state from the Recognition module.
  
• Remember the information, by keeping track of the user's emotional responses via storage in short, medium, and long-term memory buffers.
  
• Model the user's current mood, by detecting meta-patterns in the user's emotional responses over time, comparing these patterns to the user's previously-defined moods, and possibly canonical, universal or archetypal definitions of human moods.
  
• Model the user's emotional life. Recognize patterns in the way the user's emotional states may change over time, to generate a model of the user's emotional states--patterns in the typical types of emotional state that the user experiences, mood variation, degrees of valence (i.e. mildly put off vs. enraged), pattern combinations (i.e. tendencies toward a pattern of anger followed by depression).
  
• Apply the user affect model. This model may help the Affective Understanding module by informing the actions that this module decides to take--actions that use the Applications and Interface modules to customize the interaction between user and system, to predict user responses to system behavior, and to eventually make predictions about the user's interaction with environmental stimuli. The Affective Understanding module's actions may be in the form of selecting an application to open, for example an application which might interactively query the user regarding an emotional state. Or the module might supply such an application with needed information, for instance enabling the application to offer the user assistance that the user might want or need. Alternatively, the Affective Understanding module might automatically respond with a pre-user-approved action for the given situation, such as playing uplifting music when the user is feeling depressed. The Affective Understanding module might even open an application that can carry on a therapeutically-motivated conversation with its user via a discrete speech interface, perhaps discussing the user's feelings with the user after sensing that the user is alone and extremely upset.
  
• Update the user affect model. This model must be inherently dynamic in order to reflect the user's changing response patterns over time. To this end, the system will be sensitive to changes in the user's meta-patterns as it begins to receive new kinds of data from the Recognition module. Similarly, the Understanding module's learning agents will receive feedback from both Application and Interface modules that will inform changes to the user model. This feedback may consist of indications of levels of user satisfaction--whether the user liked or disliked the system's behavior. This feedback may come either as direct feedback from the user via the interface, or indirectly by way of inference from how an application was used (e.g. the way that application X was used and then terminated indicated that the user may have been frustrated with it). These user responses will help to modify the Understanding module's model of the user and, therefore, the recommendations for system behavior that the Understanding module makes to the rest of the system.
  
• Build and maintain a user-editable taxonomy of user preferences, for use in specific circumstances when interacting with the user. For example, instructions not to attempt to communicate with the user while s/he is extremely agitated, or requests for specific applications during certain moods--i.e. "Start playing melancholy music when I've been depressed for x number of hours, and then start playing upbeat music after this duration." This taxonomy may be eventually incorporated into the user model; however, ultimately, the user's wishes should be able to override any modeled preference.
  
• Feature two-way communication with the system's Recognition module. Not only will the Recognition module constantly send updates to the Understanding module, but the Understanding module will also send messages to the Recognition module. These messages may include alerting the Recognition module to "look out" for subsequent emotional responses that the Understanding module's model of the user's meta-patterns predicts. Other kinds of Understanding-module-to-Recognizing-module messages may include assisting the Recognition module in fine-tuning its recognition engine by suggesting new combinations of affect response patterns that the user seems to be displaying. These novel combinations may in turn inform novel patterns in the user's affect that may be beyond the scope of the Recognition engine to find on its own.
  
• Eventually build and maintain a more complete model of the user's behavior. The more accurate a model of the user's cognitive abilities and processes can be built, the better the system will be at predicting the user's behavior and providing accurate information to the other modules within the Affective Computing system. To this end, the field of Artificial Intelligence may be able to inform such a model. For example, the Affective Understanding module may be able to communicate with an outboard, independent system such as Lenat's Cyc (a vast data structure that attempts to model human cognition). The Cyc system might be able to functionally provide a model of the user's behavior, rationale, and motivations to the Affective Understanding module.
  
• Eventually model the user's context. The more information the system has about the user's outside environment, the more effective the interaction will be, as will be the benefit to the user. A system that knows that the user is in a conversation with someone else may not wish to interrupt the user to discuss the user's current affective response. Similarly, a system that can tell that the user has not slept in several days, is ill or starving or under deadline pressure, will certainly be able to communicate with much more sensitivity to the user. Other research on context modeling at the Media Lab includes the Smart Rooms project, such as the Artificial Life Interactive Video Environment (ALIVE).
  
• Provide a basis for the generation of synthetic system affect. A system that can display emotional responses of its own is a vast, distinct area of research. The Affective Understanding module described here may be able to inform the design of such systems. And, once built, such a system could be integrated into the Affective Understanding module to great effect. For example, a system that is able to display authentic empathy in its interaction with the user might prove even more effective in an Active Listening application than did our system that shows artificial empathy (looks like empathy to the user, but the machine doesn't really feel anything).
  
• Ensure confidentiality and security. The understanding module will build and maintain a working model and record of the user's emotional life; eventually, this model may also record other salient, contextual aspects of the user's life. Therefore, perhaps more so than any other part of the affective computing system, the affective understanding module will house information that must be kept confidential. Part of our policy is to place the highest priority on maintaining the user's privacy and control over the release of any such information. The Affective Understanding module must therefore employ strong security measures, which must be supported by the rest of the system.