Pontificia Universidad Católica de Chile Pontificia Universidad Católica de Chile
Pichara K., Protopapas P. and León D. (2016)


Revista : Astrophysical Journal
Volumen : 819
Número : 1
Páginas : 11pp
Tipo de publicación : ISI Ir a publicación


In the last few years there has been a dramatic increase in the research projects related to the development of automatic tools to explore astronomical databases. Scientists already have developed solutions to tackle several science problems, such as automatic detection of RR Lyrae stars, automatic detection of quasars, or automatic classification between periodic or non-periodic types of variable stars, among others. Most of the solutions represent astronomical objects using different models depending on the goal. As an example, continuous auto-regressive models are shown to be good to differentiate quasars from the rest of the stars, while some periodicity descriptors are good to detect RR Lyrae stars. These experts in specific challenges use different amount of resources, usually the hardest the problem the more resources are needed. Given that every time new science problems are appearing, it is very important to be able to re-use the models experience acquired before, without re-building everything from scratch when the science problem changes, integrating previous models that are closely related to the new challenges. In this paper we propose a new model that automatically integrates previous models for automatic classification of variable stars, being able to solve new classification challenges mixing the knowledge from previously trained experts in problems related to the new challenge. The integration is not trivial in the sense that each previous model is trained in a different context, answering different questions and using different representations of data, forbidding us to use most of mixture of experts algorithms in machine learning literature, where in general every expert work in the same context. Given that each expert in general uses different resources, we are encouraged to integrate with efficiency, using the most expensive experts only when it is necessary, and gathering models that together build for us the confidence to solve the actual challenge with higher levels of accuracy. We try our model with EROS-2, MACHO and OGLE datasets, showing that we solve most of the automatic classification challenges only training a model to learn how to integrate the previous experts, without having to learn any new variability pattern directly related to the stars.