Embeddings (CLS- oken) o es examples o
he seen-classes + UMAP ( ):
Su p isingly clean sepa a ion o TDE and
SNIa ( oo good?): his is a bi suspicious...
No ou lie s (expec ed?)
Two clus e s o TDEs: e y likely due o
he obse a ion pa e n? (2 gaps s 1)
Lea ning Ligh Cu e Embeddings wi h
Ro a y Masked Au oencode
Gab iella Con a do (Uni e si y o No a Go ica, SISSA), And eja Gomboc (UNG), U os
Zi ano ic (Uni e si y o T ies e, SISSA), Alex Razim (UNG), Edua do Concepcion
Cas o (UNG), Sap ashwa Bha acha yya (UNG)
a Xi :
2505.20535
A T ans o me o I egula ly Sampled Mul i-Va ia e Time-Se ies
Class disco e y and anomaly de ec ion in
Rubin’s ligh cu es → need o ans o m he
ligh cu es in o some ea u e space.
RoMAE: a speci ic posi ional encoding
(“ o a y”) o ime (and il e s) o i egula ly
sampled mul i- a ia e ime-se ies. S a e-o -
he-a esul s on supe ised asks.
A e RoMAE’s embeddings lea ned in a sel -
supe ised way (no label) use ul o
anomaly/class de ec ion in Rubin’s da a?
Unsupe ised aining wi h only wo classes o ELAsTiCC2
DP1 ligh -cu es a e sho e han ELASTiCC2 + di e en
samplings: could explain why hey a e well sepa a ed (+
aining wi h only 2 classes). Does no explain why one
lands in + no in he igh egion?
Conce ning beha io in ELAsTiCC: he e is such a hing
as oo good esul s...
Small “clus e s” in DP1 ha d o in e p e ; neighbo s o
SN om F eebu n don’ eally look like SN... No clea
sepa a ion in TDE/SNIa using ZTF.
UMAP no ideal: u u e wo k ain o smalle
ep esen a ions di ec ly; ain wi h mo e classes/ a ie y.
Injec ing New Classes + Rubin DP1-ECDFS
Real da a
Ligh -cu es o DP1-ECDFS a e cons uc ed om DiaObjec s and DiaSou ces wi h all lagged
obse a ions emo ed, keeping objec s wi h a leas 3 de ec ions in 2 il e s
T ain on DP1 only T ain on ZTF TDE
& SNIa only Some hough s
Acco ding o he models, DP1-ECDFS ligh -
cu es look e y di e en om he ELAsTiCC2
classes seen in aining (excep his one )
Case 1:
TDE+EB
Case 2:
TDE+SNIa
