derek ruths || network dynamics

Understanding exercising behavior

Exercise forms a central part of many peoples’ healthy lifestyles with individuals choosing different activities to fit their interests and time availability. These exercising habits often follow regular trends. For example, we could roughly categorize groups of people as “runners”, “crossfitters”, or “weightlifters” based on the activities they perform. These types of exercising behaviors can reveal emerging trends in how people seek to improve their health and, when we look at how these behaviors differ across age and gender, and can tell us how people evolve in their exercising habits.

In our ICWSM 2015 study, we examined the complete exercising history of over 440,000 individuals on the website Fitocracy to analyze the underlying patterns of exercising behavior. Notably, Fitocracy allows its users to provide their age and gender, which lets us peek into how different ages and genders behave. In order to identify the underlying behaviors from the activities, we trained a Latent Dirichlet Allocation (LDA) model on users’ exercising data. In our setting, we treat each individual’s history as a reflection of that person engaging in just a few behaviors, where each behavior selects for specific exercises with higher frequency. For example a triathlete’s workout history might be a mixture of “running,” “swimming,” and “cycling” behaviors, where the “running” behavior might be shown in the individual doing exercises like “trail running”, “sprints”, “jogging”, and so on. The beauty of the LDA model is that we as experimenters don’t need to manually decide on which behaviors people have; instead, the model learns these behaviors automatically from the data itself to tell us what people are truly doing!

The results in the study’s paper highlighted several interesting behaviors learned when the model was asked to learn 20 behaviors. Due to space constraints we were unable then to show all of the interesting behaviors and therefore, we present these results here as an interactive demonstration. The visualization allows you to choose how many behaviors the model was asked to find (e.g., 20) and then shows the populations engaging in the behavior. On the left, you can see a table with the demographic summary per behavior. Click on any behavior to view its population demographics in the center. You can also use the arrow keys to scroll through the behaviors too! On the right you will see the exercises that people with this behavior are most likely to engage in. For example, Behavior 1 for the 20-behavior study is most associated with stretching and warm-up activities, which are more likely to be done by females than males!

Please feel free to explore the different behaviors and see how different ages and genders exercise. You may be surprised! For full details on how this model was created, please see our paper at ICWSM or contact the first author David Jurgens.

To interact with this chart:

  • Press the up and down arrow keys or ‘n’ and ‘b’ to move up and down the behavior list.
  • Click on any Behavior’s row in the left table to view that behavior in the center screen.
  • The columns of the left table are sortable; just click on the column header that you want to sort.
  • Select from the “Number of Behaviors” menu to choose how many behaviors were learned by our LDA model.

Fitocracy Exercise Behaviors