As a Researcher I want to infer the home location from a trajectory

[noerw]

• specify semantics of sparse input trajectory data:

  • How do we differentiate between "time between points is time stayed at that point" and "time between points encodes missing spatial data"?
    I'm sure there already plenty of literature & code on this topic.
    A simple and (likely efficient) approach would be to annotate the first point of a new segment / recording session in the databaseare periods between points considered to mean "did not move", or NA, or do we apply some uncertainty metric?

    • Of the example trajectories collected, most consist of several segments (i.e. regular segments combining into a sparse trajectory), so it would make sense to store them as MultiLineStrings.
    • For the user generated trajectectory, location updates are temporally sparse all the time, so it's hard to make a cut between segments in order to interpret temporal gaps in the data.

  • ...?

• conceptualize an inference engine architecture. ideas:

  • modular: plug in new inference matchers
  • classes of inference matchers, parametrized? for example

    homeInference = new MedianLocationInference({ filters: { timeofDay: [23, 6] } })
    workInference = new MedianLocationInference({ filters: { timeofDay: [10, 14], dayOfWeek: [0, 5] } })

  • streaming: inferences ideally should be able to be computed live on the stream of incoming points, to avoid recomputation on the whole trajectory for each point. This requires some form of serialization/persistence of the inference model state.
  • not relying on DB queries, for it to run in web environment as well? (can we deal with the performance?)