Deep classifier · clf-v2026.06.17

How the classifier is graded

Every star carries a predicted class from a deep neural network. Here is how well it does, measured on the held-out test split it never trained on — the scorecard, confusion matrix and all.

Fine accuracy

90.1%

21 classes · 39,530 stars

Fine macro-F1

0.819

unweighted over classes

Broad accuracy

98.0%

8 families · rollup

Broad macro-F1

0.966

balanced acc 0.960

clf-v2026.06.17 held-out test split · 39,530 labeled

Where it confuses classes

Confusion matrix

Each row is a true class; each cell shows what fraction the model predicted as the column class. A bright diagonal is the goal; the off-diagonal glow is where look-alikes blur — EW↔EB, the RR Lyrae subtypes, semiregulars↔irregulars. Hover any cell for the counts.

0% 100% of row diagonal = correct · off-diagonal = confusion

Class by class

Precision, recall & F1

Precision is how often a predicted label is right; recall is how many of the true members it finds. Rare classes with little training data (RRd, type-II Cepheids) sit lower.

Broad · 8 classes

precision · recall · F1 · support

class	precision	recall	F1	support
ECLIPSING	0.985	1.000	0.993	9,800
RR_LYRAE	1.000	1.000	1.000	3,880
CEPHEID	1.000	0.978	0.989	1,020
DELTA_SCUTI	1.000	1.000	1.000	2,060
LPV	0.988	1.000	0.994	14,600
ROTATIONAL	0.943	0.930	0.936	4,200
ERUPTIVE_CV	0.886	0.886	0.886	1,290
OTHER	0.974	0.886	0.928	2,680

class	precision	recall	F1	support
EW	0.900	0.960	0.929	5,200
EA	0.886	0.900	0.893	3,100
EB	0.793	0.660	0.720	1,500
RRAB	0.959	0.970	0.965	2,600
RRC	0.891	0.880	0.886	1,100
RRD	0.600	0.550	0.574	180
DCEP	0.854	0.919	0.885	520
DCEPS	0.825	0.660	0.733	150
CWA	0.802	0.742	0.771	240
CWB	0.674	0.582	0.624	110
DSCT	0.942	0.910	0.926	1,800
HADS	0.495	0.612	0.547	260
M	0.964	0.970	0.967	6,400
SR	0.856	0.920	0.887	5,800
L	0.890	0.780	0.831	2,400
ROT	0.943	0.930	0.936	4,200
YSO	0.677	0.741	0.708	640
GCAS	0.629	0.641	0.635	220
CV	1.000	0.851	0.920	430
AGN	1.000	0.960	0.980	980
OTHER	0.931	0.820	0.872	1,700

Beyond the headline

Slices & diagnostics

asassn discovery

91.8%

macro-F1 0.741 · 41,000 stars

vsx native

93.4%

macro-F1 0.762 · -1,470 stars

Blazhko anomaly AUC

0.680

does the anomaly score rank 214 Blazhko RRab above ordinary RRab? (> 0.5 = yes)

Ambiguous-family accuracy

87.1%

of 7,400 stars VSX could only type loosely (E, RR, CEP…), landing somewhere in the right family

How to read this

The classifier predicts 21 fine classes rolled up into 8 broad families; broad probabilities are sums of fine ones, so the two levels never disagree. Training excludes labeler-uncertainty codes and suspected variables, and ASAS-SN's own discoveries are scored as a separate slice (their types trace back to an earlier classifier). Every number here is measured on stars held out of training — never the data the model saw. Disagree with a label? The class probabilities, anomaly scores, and the full light curve are on each star page, and you can re-run the classifier on your own curve in the analyzer.