RAMP 03 Astrophyiscs

vendredi 10 avr. 2015, 09:00 → 18:00 Europe/Paris

Proto204

Akin Kazakci (MINES ParisTech), Alexandre Gramfort (Telecom ParisTech, CNRS), Balázs Kégl (LAL), Djalel Benbouzid (LAL, Université Paris-Sud), Loïc Estève (INRIA)

This RAMP will be on Astrophysics, more precisely, on classification of variable stars from their light curves (luminosity vs time profiles), brought to you by Marc Monier (LAL), Gilles Faÿ (Centrale-Supelec), and your regular coaches.

The event will take place at Proto204, a 10 minute walk from either the Orsay-Ville or the Bur-sur-Yvette RER B stations.

Supporting material:

• static table
• time series (light curves)
• ipython notebook with the problem description, analysis examples, and submission instructions
• example of classifier.py
• example of feature_extractor.py
• user_test_model.py: unit test for pre-checking the submission: run
  python user_test_model.py
  before submitting.

Some general guidelines. 

Your primary goal is to have a high score in the "contributivity" column. One way to achieve it is to submit a strong model, high in the "score" column, but we appreciate especially those models which do not have top scores but are sufficiently different from the rest of the models to achieve high score in the "contributivity" column.

Our best feature extractor is based on Gaussian Process. It smoothes the light curves, aligns them so the period starts at the minimum luminosity, normalizes the amplitude so the response is bounded between 0 and 1, and bins the smooth curve into ten bins. It also adds the amplitude and the lenght scale (smoothness) as feature. Here are some plots that show how it works.

The performance gain is significant, so you might want to work on this model if you decide to improve the feature extractor. The bad news is that it is quite slow: it takes eight minutes to train. So, if you submit anything based on this feature extractor, please do extensive preliminary tests on your computer or on your private VM, and do not submit more than once per 30 minutes to an hour. A general rule is that you should not resubmit until your previous model has not yet been trained (it shows up in the "New models" table.)

Some strategies you can employ:

1. Take a good classifier and work on the feature extractor or vice verse.
2. Take a good classifier or feature extractor and optimize its hyperparameters using smac or hyperopt. See the course material for help.
3. Take two good submissions and combine the feature extractor and the classifier.

Welcome of participants & Coffee

Introduction talk

Data munging and machine learning session 1

Buffet & presentation of the first results

Data munging and machine learning session 2

Café

Data munging and machine learning session 3

Debriefing and closing