Abstract:In the paper we develop multi-class classification of Fermi-LAT gamma-ray sources using machine learning with hierarchical determination of classes. One of the main challenges in the multi-class classification of the Fermi-LAT sources is that the size of some of the classes is relatively small, for example with less than 10 associated sources belonging to a class. In the paper we propose a hierarchical structure for the determination of the classes. This enables us to have control over the size of classes and to compare the performance of the classification for different numbers of classes. In particular, the class probabilities in the two-class case can be computed either directly by the two-class classification or by summing probabilities of children classes in multi-class classification. We find that the classifications with few large classes have comparable performance with classifications with many smaller classes. Thus, on the one hand, the few-class classification can be recovered by summing probabilities of classification with more classes while, on the other hand, the classification with many classes gives a more detailed information about the physical nature of the sources. As a result of this work, we construct three probabilistic catalogs. Two catalogs are based on Gaussian mixture model for the determination of the groups of classes: one with the requirement for the minimal number of sources in a group to be larger than 100, which results in six final groups of physical classes, while the other one has the requirement that the minimal number of sources in a group is larger than 15, which results in nine groups of physical classes. The third catalog is based on a random forest determination of the groups of classes with the requirement that the minimal number of sources in a group is larger than 100, which results in six groups of classes. (abridged)

Effect of covariate shift on multi-class classification of Fermi-LAT sources

Classification of Fermi-LAT unassociated sources with machine learning in the presence of dataset shifts

Multi-class classification of Fermi-LAT sources with hierarchical class definition

Towards probabilistic multiclass classification of gamma-ray sources

Beyond Reweighting: On the Predictive Role of Covariate Shift in Effect Generalization

A Statistical Approach to Recognizing Source Classes for Unassociated Sources in the First Fermi-LAT Catalog

Harnessing the Power of Vicinity-Informed Analysis for Classification under Covariate Shift

Covariate-Shift Generalization Via Random Sample Weighting.

Double-Weighting for Covariate Shift Adaptation

Federated Learning under Covariate Shifts with Generalization Guarantees

Classification under Nuisance Parameters and Generalized Label Shift in Likelihood-Free Inference

Model-free selective inference under covariate shift via weighted conformal p-values

Maximum Likelihood Estimation is All You Need for Well-Specified Covariate Shift

Nearest Neighbor Sampling for Covariate Shift Adaptation

A One-step Approach to Covariate Shift Adaptation

Robust Covariate Shift Adaptation for Density-Ratio Estimation

Covariate shift in nonparametric regression with Markovian design

Transfer Learning under Covariate Shift: Local $k$-Nearest Neighbours Regression with Heavy-Tailed Design

Double Debiased Covariate Shift Adaptation Robust to Density-Ratio Estimation

Algorithmic Fairness Generalization under Covariate and Dependence Shifts Simultaneously

Estimation of redshift and associated uncertainty of Fermi/LAT extra-galactic sources with Deep Learning