CFP: Let’s predict your activity for next week

Last week, we created a Critical Experience Prototype (CEP) that tested whether wearable users appreciate a more personalized experience with their device and application. This week, we created a Critical Function Prototype (CFP) that tests whether the most essential function of a possible solution can be realized with today’s technology.

But what are we going to implement? As we were puzzled at the beginning, we did a traditional brainstorm session, in which we came up with the following ideas:

  • Integrate all wearable data into one data hub
  • Do predictions based on wearable data
  • Combine wearable data with other kind of personal data (e.g. calendar)

We liked the ideas, however none of them were very novel. But the combination of the three was a unique thought.

So our idea was to integrate all wearable data into one central data hub, combine that data with calendar data based on which the activity level of your upcoming weeks is predicted. But how does it work?

Busy people have a full calendar with frequently recurring events. For instance, a programmer might have the daily scrum meeting at work. During these events people usually walk the same number of steps. If we can classify future calendar events as a recurring event, we can predict the estimated number of steps the person will walk during the event. If we can classify all future events of a person, we can give a good estimate about the upcoming activity level.

The prototype we implemented used one of our member’s wearable and calendar data. The wearable data had a granularity of 1 hour and the calendar was completely filled. To simplify the implementation, we assumed a normal distribution of steps during measurement intervals. Furthermore, the location was excluded from the calendar.

The combination of the two data sources intersects the time intervals of the calendar data with the wearable data. With the intersected time intervals, we calculate the average steps per hour for the event. The average steps enable us to estimate the number of steps for future events and thus to predict the future activity level.

The result looks as follows:

It shows that we still have lots of variance in terms of step count per hour for the events. This is due to the fact that we only used one week of calendar and wearable data. The more data we use, the more accurate will be our model. However, the diagram shows that each event has different step counts per hour and correctly classifies that during soccer our team member walked the most. Therefore, we can conclude that it is possible to combine calendar with wearable data despite its apparent shortcomings.

What is next? We will further elaborate on this idea and we already have some ideas on how to think it further:

  • Integrate more wearable and calendar data into the model
  • Extend the model to work with calendars that miss events
  • Extend the model to work with location data
  • Make the prediction motivate the user to beat the prediction