Abstract
Abstract
We present an uncertainty-aware, fine-grained multiclass classification of unresolved Fermi Large Area Telescope (LAT) gamma-ray sources in the 4FGL-DR4 catalog. We define the unresolved-source population as blank-class sources, sources labeled “
UNK
,” and blazar candidates of uncertain type, and classify them within a supervised 22-class catalog taxonomy. The baseline analysis represents each source through the three catalog spectral forms—power law (PL), log parabola (LP), and PL with a superexponential cutoff (PLEC)—and compares TabNet, random forest, and a stacked TabNet + random forest ensemble. The ensemble reaches accuracies of 92.95%, 92.45%, and 94.13% for the PL, LP, and PLEC representations, respectively. To quantify the role of feature choice, we further perform controlled ablations using an all-spectral configuration, which combines PL, LP, and PLEC parameters simultaneously, together with LP-only, PL-only, PLEC-only, and photometry + variability representations. Each configuration is tested with and without uncertainty columns and with
λ
max
=
0.1
–0.5 for synthetic interpolation. The strongest ablation result is obtained by the uncertainty-aware PLEC configuration, with accuracy and weighted
F
1
of 0.9614. We also provide probabilistic predictions for all 4186 unresolved sources, including top-3 class probabilities, full class-probability vectors, top 1 – top 2 probability margins, and predictive entropy.