Discussion on "Exploiting Non-Linear Structure in Astronomical Data for Improved Statistical Inference" by Ann B. Lee and Peter E. Freeman

Didier Fraix-Burnet (1)

1 IPAG - Institut de Planétologie et d'Astrophysique de Grenoble

Both dimensionality reduction and classification seek a reduced simpler form of the data. The first one works with the parameter space, while classification works with the object space. Ideally, one wishes to find a parameter space in which the points are naturally gathered into distinct groups and, as a physicist more particularly, data points can fit our model curves. I want to point out that dimensionality reduction methods and classification approaches are highly complementary and should even be carried out together. Astrophysical objects are complex, so that numerical simulations are now a common tools to do physics. Model fitting has thus become a comparison between populations (the observed ones and the synthetic ones) rather than plotting a curve onto data points. This is exactly the role of statistics.

Mots clés: Classification; Dimensionality Reduction; Populations

Domaines: Applications [stat.AP]; Instrumentation et méthodes pour l'astrophysique [astro-ph.IM]; Instrumentation et méthodes pour l'astrophysique [astro-ph.IM]

Didier Fraix-Burnet. Discussion on "Exploiting Non-Linear Structure in Astronomical Data for Improved Statistical Inference" by Ann B. Lee and Peter E. Freeman. Statistical Challenges in Modern Astronomy V, Jun 2011, State College (PA), United States. ⟨hal-00627077⟩