Behavior Dashboard is a visual analytics tool that allows users to explore and interact with computational models for human routine behavior.
Behavior Dashboard provides an explorative means for users to understand the routine behaviors of people under various circumstances and domains. Human routine behaviors can be modeled as a Markov decision process comprised of a sequence of situations and actions: a situation is the current state that a person might be in, while an action is a transition from one situation to another. Situation features are represented as rows of orange cells while action features are represented as rows of blue cells. Hovering over a cell displays a tooltip that shows the feature, the name of the feature value being hovered over, and the probability that feature value is satisfied.
A behavior instance is a sequence of situations and actions. A routine variation is a combination of related behavior instances. Users can explore behavior instances and hypothetical routine variations through Data Tracks and Constraint Tracks, respectively.
With Data Tracks, users can specify ‘estimated features’, which are features the model will try to predict. Specifying the estimated features for a data track allows users to evaluate the accuracy of the model by comparing model predictions against the actual values of specific features.
One of the strengths of Behavior Dashboard is that it allows users to see how the behavior of a population changes over time. In doing so, they may be able to confirm hypotheses about whether certain behaviors can lead to specific outcomes.
So far, we have used Behavior Dashboard to explore the relationship between the behaviors of pancreatic surgery patients and hospital readmission over time. We found that an increase in physical activity does not necessarily predict improvement in pain or nausea symptoms over time, which suggests that a patient’s ability to perform physical activity is more likely to be affected by their symptoms instead of the other way around. Instead of exclusively promoting light physical activity, we believe any behavior-based intervention that tries to alleviate readmission should also help manage patients’ symptoms.
I built the Behavior Dashboard interface using Java and Processing. All the interactive elements were made from scratch, with the exception of the buttons/toggles/dropdown menus which were created with the help of controlp5. The Dashboard relies on behavior models developed by Nikola Banovic. Each time the interface has to load new data, it must communicate with the computational model in the backend, which often contains millions of possible transitions between situations and trillions of possible behavior instances. Angela Xie was responsible for the lovely visual design, you can read about her design process here.
Nikola Banovic, Tofi Buzali, Fanny Chevalier, Jennifer Mankoff, and Anind K. Dey. 2016. Modeling and Understanding Human Routine Behavior. In Proceedings of the 2016 CHI Conference on Human Factors in Computing Systems (CHI '16). ACM, New York, NY, USA, 248-260. DOI: https://doi.org/10.1145/2858036.2858557
Nikola Banovic, Anqi Wang, Yanfeng Jin, Christie Chang, Julian Ramos, Anind Dey, and Jennifer Mankoff. 2017. Leveraging Human Routine Models to Detect and Generate Human Behaviors. In Proceedings of the 2017 CHI Conference on Human Factors in Computing Systems (CHI '17). ACM, New York, NY, USA, 6683-6694. DOI: https://doi.org/10.1145/3025453.3025571
This project is a work-in-progress. Stay tuned for updates!