P.17 Gabriel Molina (Universidad Técnica Federico Santa María)
Can we interpret machine learning? An analysis of exoplanet detection problem.
The exoplanet detection problem - planets that orbit a star outside our Solar System - has focussed on the use of time-consuming manual process. Now, the promising techniques are machine learning methods. However, the lack of interpretability in order to understand what the models does, has avoided the improvement and development of the models. In this work, we study the use of classical machine learning methods for detecting confirmed objects on the Kepler mission. As previous works have explored, we used metadata from the objects and hand-crafted features from the light curves. Our study shows that approximately 93% of the data is correctly detected. We also show that the behavior of non-exoplanet objects is quite extreme on some features, which facilitate the recovery of mostly all these objects (high recall). At the same time, our study presents difficulties with confirmed objects, where the behavior that the features express overlap with the non-exoplanet objects, leading to a contaminate prediction of exoplanet objects (low precision). Note that our work presents some insights about where the error is. For example, earth-like planets (small and relatively far apart from their mother star) have not been able to study due to physical limitations.
