feature extraction from irregularly-spaced timeseries

This is just getting started, and the API is going to change a lot.

The motivation for this is a clinical data mining application.

The idea is to have a very compact representation of a series of events, where each event has:

• a timestamp (uint32)
• a "subject" (int32)
• a "verb" (int32)
• an optional value (float32)

In our clinical data analysis application, the "subject" is a patient, the "verb" is an event that happens, such as admission, medication prescribed, or a particular lab test result, and the "value" is an optional value associated with the event, such as the numerical lab result value. We have hundreds of millions of events, and we want to work with them in memory on a single node.

Given a series of events, at any point in time we want to be able to build a fixed-length feature vector that reflects the state of a particular subject at that time. This vector should summarize the preceeding events, and be usable as features in a machine learning model. For example, one element in that vector might be the number of events of a particular verb that subject has had up to that point. Another element might be the value for the most recent event of a certain verb. Most of this is not implemented yet.

Install:

cd eventlearn
pip install .

Run tests:

py.test tests