MNRAS 489, 3591–3608 (2019) doi:10.

1093/mnras/stz2362
Advance Access publication 2019 August 26

Anomaly detection in the Open Supernova Catalog

M. V. Pruzhinskaya,1‹ K. L. Malanchev ,1,2‹ M. V. Kornilov,1,2 E. E. O. Ishida,3

F. Mondon,3 A. A. Volnova4 and V. S. Korolev5,6

Downloaded from https://academic.oup.com/mnras/article/489/3/3591/5554761 by Universidad Nacional Autonoma de Mexico user on 04 June 2024
1 Lomonosov Moscow State University, Sternberg Astronomical Institute, Universitetsky pr. 13, Moscow, 119234, Russia
2 National Research University Higher School of Economics, 21/4 Staraya Basmannaya Ulitsa, Moscow, 105066, Russia
3 Université Clermont Auvergne, CNRS/IN2P3, LPC, F-63000 Clermont-Ferrand, France
4 Space Research Institute of the Russian Academy of Sciences (IKI), 84/32 Profsoyuznaya Street, Moscow, 117997, Russia
5 Central Aerohydrodynamic Institute, 1 Zhukovsky st, Zhukovsky, Moscow Region, 140180, Russia
6 Moscow Institute of Physics and Technology, 9 Institutskiy per., Dolgoprudny, Moscow Region, 141701, Russia

Accepted 2019 August 21. Received 2019 August 15; in original form 2019 March 24

ABSTRACT
In the upcoming decade, large astronomical surveys will discover millions of transients raising
unprecedented data challenges in the process. Only the use of the machine learning algorithms
can process such large data volumes. Most of the discovered transients will belong to the
known classes of astronomical objects. However, it is expected that some transients will be
rare or completely new events of unknown physical nature. The task of finding them can
be framed as an anomaly detection problem. In this work, we perform for the first time an
automated anomaly detection analysis in the photometric data of the Open Supernova Catalog
(OSC), which serves as a proof of concept for the applicability of these methods to future large-
scale surveys. The analysis consists of the following steps: (1) data selection from the OSC
and approximation of the pre-processed data with Gaussian processes, (2) dimensionality
reduction, (3) searching for outliers with the use of the isolation forest algorithm, and (4)
expert analysis of the identified outliers. The pipeline returned 81 candidate anomalies, 27
(33 per cent) of which were confirmed to be from astrophysically peculiar objects. Found
anomalies correspond to a selected sample of 1.4 per cent of the initial automatically identified
data sample of approximately 2000 objects. Among the identified outliers we recognized
superluminous supernovae, non-classical Type Ia supernovae, unusual Type II supernovae, one
active galactic nucleus and one binary microlensing event. We also found that 16 anomalies
classified as supernovae in the literature are likely to be quasars or stars. Our proposed pipeline
represents an effective strategy to guarantee we shall not overlook exciting new science hidden
in the data we fought so hard to acquire. All code and products of this investigation are made
publicly available.1
Key words: methods: data analysis – catalogues – supernovae: general.

study of different types of SNe allows us to probe the composition

1 I N T RO D U C T I O N
and distance scale of the Universe (Kirshner & Kwan 1974; Riess
Supernovae (SNe) hold vital pieces of the large cosmic puzzle et al. 1998; Perlmutter et al. 1999; Hamuy & Pinto 2002) – imposing
astronomy and cosmology aim to solve. They are responsible for the strong constraints on the standard cosmological model (Betoule
chemical enrichment of interstellar medium (Nomoto, Kobayashi & et al. 2014; Scolnic et al. 2018).
Tominaga 2013); the production of high-energy cosmic rays (Mor- Given the potential impact of SN research on different areas of
lino 2017), and they trigger star formation via the density waves astronomy, the scientific community has allocated a large fraction
induced by their energetic explosions (Nagakura, Hosokawa & of its efforts in the generation of large SN surveys – a few recent
Omukai 2009; Chiaki, Yoshida & Kitayama 2013). Moreover, the examples include the Carnegie Supernova Project2 (CSP; Hamuy
et al. 2006), the Panoramic Survey Telescope and Rapid Response

 E-mail: [email protected] (MVP); [email protected]

(KLM)
1 http://snad.space/osc/ 2 https://csp.obs.carnegiescience.edu/


C 2019 The Author(s)

Published by Oxford University Press on behalf of the Royal Astronomical Society

3592 M. V. Pruzhinskaya et al.
System3 (Pan-STARRS; Kaiser et al. 2010; Chambers et al. 2016), maximum brightness that is not consistent with radioactive decay
the Dark Energy Survey4 (DES; Dark Energy Survey Collaboration of 56 Ni (Ben-Ami et al. 2014), ASASSN-15lh – for some time it
et al. 2016), and the Zwicky Transient Facility5 (ZTF; Bellm et al. was considered as the most luminous SN ever observed – two times
2019). Another generation of even larger counterparts, like the Large brighter than superluminous SNe (SLSN), later the origin of this
Synoptic Survey Telescope6 (LSST; LSST Science Collaboration object was challenged and now it is considered as a tidal disruption
et al. 2009), will soon join this list, making available a combined of a main-sequence star by a black hole (Dong et al. 2016; Leloudas
data set of unprecedented volume and complexity. et al. 2016). As such sources are typically rare, the task of finding

Downloaded from https://academic.oup.com/mnras/article/489/3/3591/5554761 by Universidad Nacional Autonoma de Mexico user on 04 June 2024
In this new data paradigm, the use of machine learning (ML) them can be framed as an anomaly detection problem.
methods is unavoidable (Ball & Brunner 2010). Astronomers In this paper, we turn to the automatic search for anoma-
have already benefited from developments in machine learning, in lies in the real photometric data using the Open Supernova
particular for exoplanet search (McCauliff et al. 2015; Thompson Catalog7 (OSC; Guillochon et al. 2017). The OSC has never
et al. 2015; Pearson, Palafox & Griffith 2018), but the synergy been used for the task of the anomaly detection with the ML
is far from that achieved by other endeavours in genetics (Chen & algorithms until this work, however, it was used for the classification
Ishwaran 2012; Libbrecht & Noble 2015; Quang & Xie 2016), ecol- problem (Narayan et al. 2018; Muthukrishna et al. 2019). The
ogy (Criscia, Ghattasb & Pererac 2012), or medicine (Dubost et al. anomalies we are looking for are any artefacts in the data, cases
2019; Venkatraghavan et al. 2019). Moreover, given the relatively of misclassification [active galactic nuclei (AGNs), novae, binary
recent advent of large data sets, most of the ML efforts in astronomy microlensing events], rare classes of objects (SLSN, kilonovae, SNe
are concentrated in classification (e.g. Kessler et al. 2010; Ishida & associated with gamma-ray bursts), and objects of unknown nature.
de Souza 2013; Heinis et al. 2016; Lochner et al. 2016; Sooknunan We use the isolation forest as an outlier detection algorithm that
et al. 2018; Ishida et al. 2019) and regression (e.g. Hildebrandt identifies outliers instead of normal observations (Liu, Ting & Zhou
et al. 2010; Cavuoti et al. 2015;Beck et al. 2017; Vilalta et al. 2017) 2012). This technique is based on the fact that outliers are data points
tasks. Machine learning is also actively applied for the real-bogus that are few and different. Similarly to random forest it is built on
classification that allows to automatically disentangle real transients an ensemble of binary (isolation) trees. The final goal of this work
from the artefacts on the images produced by major time-domain is to develop some approach that allows to detect anomalies in
surveys (Bloom et al. 2012; du Buisson et al. 2015; Goldstein huge amount of data produced by time-domain surveys such as
et al. 2015; Wright et al. 2015). A large variety of ML methods LSST. Due to the initial absence of any labelled data in transient
were applied to supervised photometric SN classification problem data bases, the algorithm follows the paradigm of unsupervised
(Richards et al. 2012; Sanders, Betancourt & Soderberg 2015b; learning. For this reason, we pretend that we do not have any labels
Lochner et al. 2016; Möller et al. 2016; Charnock & Moss 2017; in the OSC and we use only the multicolour photometry. Moreover,
Revsbech, Trotta & van Dyk 2018; Brunel et al. 2019; Möller & de the spectral classification provided by the OSC is collected from
Boissière 2019; Pasquet et al. 2019) and unsupervised characteriza- different sources, including the preliminary classification from the
tion from spectroscopic observation (e.g. Rubin & Gal-Yam 2016; astronomical telegrams8 where it can be based on one spectrum
Sasdelli et al. 2016; Muthukrishna, Parkinson & Tucker 2019). only, that is simply fitted by SNID (Blondin & Tonry 2007) to
Astronomical anomaly detection has not been yet fully imple- the closest SN template. Such rough classification cannot give an
mented in the enormous amount of data that has been gathered. information about peculiar behaviour of the source, usually more
Barring a few exceptions, most of the previous studies can be divided detailed study is needed. On the contrary, it is not necessary that all
into only two different trends: clustering (e.g. Rebbapragada et al. outliers found by machine are real anomalies. That is why we also
2009) and subspace analysis (e.g. Henrion et al. 2013) methods. subject the outliers to the careful astrophysical analysis using the
More recently, random forest algorithms have been extensively used publicly available information.
by themselves (Baron & Poznanski 2017) or in hybrid statistical The rest of the paper is organized as follows. In Section 2, we
analysis (Nun et al. 2014). Although all of this has been done to describe the data used for the analysis. Section 3 is devoted to
periodic variables there is not much done for transients and even work related to the data pre-processing, including LC approximation
less for SNe. by Gaussian processes (GP). The outlier detection algorithm is
The lack of spectroscopic support causes the large SN data presented in Section 4. Section 5 shows the results and contains the
bases to collect SN candidates basing on the secondary indicators analysis of found outliers. We conclude the paper in Section 6. The
[proximity to the galaxy, arise/decline rate on a light curve (LC), outliers are listed in Appendix A.
absolute magnitude]. This leads to the appearance of incorrectly
classified objects. Anomaly detection can help us to purify the
2 T H E O P E N S U P E R N OVA C ATA L O G
SN data bases from the non-SN contamination. It is also expected
that during such analysis the unknown variable objects or SNe The data are drawn from the OSC (Guillochon et al. 2017). The
with unusual properties can be detected. As an example of unique catalogue is constructed by combining many publicly available
objects one can refer to SN2006jc – SN with very strong but data sources such as the Asiago Supernova Catalog (Barbon
relatively narrow He I lines in early spectra (around 30 similar et al. 1999), the Gaia Photometric Science Alerts (Wyrzykowski
objects are known; Pastorello et al. 2016), SN2005bf – SN attributed et al. 2012; Campbell et al. 2014), the Nearby Supernova
to SN Ib but with two broad maxima on LCs (Folatelli et al. 2006), Factory (Aldering et al. 2002), Pan-STARRS (Kaiser et al. 2010;
SN2010mb – unusual SN Ic with very low decline rate after the Chambers et al. 2016), the SDSS Supernova Survey (Sako et al.
2018), the Sternberg Astronomical Institute Supernova Light Curve
Catalogue (Tsvetkov, Pavlyuk & Bartunov 2005), the Supernova
3 https://panstarrs.stsci.edu/
4 https://www.darkenergysurvey.org/
5 https://www.ztf.caltech.edu/ 7 https://sne.space/
6 https://www.lsst.org/ 8 http://www.astronomerstelegram.org/

MNRAS 489, 3591–3608 (2019)

 SNAD: SuperNova Anomaly Detection 3593

Downloaded from https://academic.oup.com/mnras/article/489/3/3591/5554761 by Universidad Nacional Autonoma de Mexico user on 04 June 2024
Figure 1. Workflow for the analysis. Ni denotes the number of observations in i’th band. GP photometry includes 364 features: 121 × 3 normalized fluxes
and the LC flux maximum; GP parameters are nine fitted parameters of the Gaussian process kernel and the log-likelihood of the fit.

Legacy Survey (SNLS, Pritchet & SNLS Collaboration 2005; are described in detail in the subsections below and illustrated by
Astier et al. 2006), the MASTER Global Robotic Net (Lipunov et al. Fig. 1. First, we prepared the photometric data extracted from the
2010), the All-Sky Automated Survey for Supernovae (ASAS-SN, OSC; we transformed the magnitudes to the flux units, converted the
Holoien et al. 2019), and the intermediate Palomar Transient upper limits, and implemented 1-d time-binning. Then, we used the
Factory (iPTF, Law et al. 2009; Cao, Nugent & Kasliwal 2016a) Gaussian processes to approximate the photometric observations in
among others, as well as from individual publications. It represents each filter. The objects with bad LC approximations were removed
an open repository for SN metadata, LCs, and spectra in an from the further analysis. After that, we transformed the remaining
easily downloadable format. This catalogue also includes some LCs in BRI filters to gri. To have a homogeneous input data, for
contamination from non-SN objects. each object we extracted its photometry in the range [−20, +100]
Given the large number of objects and their diverse characteris- d relative to the maximum flux. We also kept the kernel parameters
tics, this catalogue is ideal for our goal of automatically identifying of the Gaussian processes. All of this together was subjected to the
anomalies. It incorporates data for more than 5 × 104 SNe dimensionality reduction procedure using t-SNE method (Maaten &
candidates among which 1.2 × 104 objects have >10 photometric Hinton 2008).
observations and 5 × 103 have spectra. For comparison, SDSS SN
catalogue contains only 4607 SNe candidates: 889 with measured
spectra (Sako et al. 2018). 3.1 Filter transformation
The catalogue stores the data in different photometric passbands. In order to ensure maximum exploitation of the data at hand, we
To have a more homogeneous sample, we chose only those objects convert the Bessel’s BRI into gri filters using the transformation
  
that have LCs in BRI (Bessell 1990), g r i or gri filters. The primed equations.10 These equations are derived by matching SDSS DR4
    
system u g r i z is defined in the natural system of the USNO 1- photometry to Peter Stetson’s published photometry for stars:11
m telescope. The SDSS magnitudes ugriz, however, are defined ⎧
in the natural system of the SDSS 2.5-m telescope. These two ⎪
⎪ B = u − 0.8116 (u − g) + 0.1313
⎪
⎪
systems are very similar and the coefficients of the transformation ⎪
⎪ B = g + 0.3130 (g − r) + 0.2271
⎪
⎪
equations are quite small (Fukugita et al. 1996; Tucker et al. 2006; ⎪
⎪
⎪
⎪ V = g − 0.2906 (u − g) + 0.0885
  
Smith et al. 2007). We assume that g r i filters are close enough ⎪
⎪
⎨ V = g − 0.5784 (g − r) − 0.0038
to gri and transform BRI to gri (see Section 3.1). We require a (1)
minimum of three photometric points in each filter with a 3-d ⎪
⎪ R = r − 0.1837 (g − r) − 0.0971
⎪
⎪
binning (Fig. 1). Our experiments show that this threshold is ⎪
⎪ R = r − 0.2936 (r − i) − 0.1439
⎪
⎪
enough to provide a good reconstruction of the LC – specially in ⎪
⎪
⎪
⎪ I = r − 1.2444 (r − i) − 0.3820
⎪
⎪
cases where photometric points are not homogeneously distributed ⎩
among filters. This is natural consequence of the LC approximation I = i − 0.3780 (i − z) − 0.3974.
procedure we adopted (Section 3.2) which takes into account the As we can see, there are several possibilities to obtain gri LCs
correlation between photometric bands to guide the reconstruction from the Bessel’s ones. Obviously, the accuracy of the transfor-
in sparsely populated filters. After this first cut, our sample consists mation increases with the number of available filters. However,
  
of 3197 objects (2026 objects in g r i , 767 objects in gri, and 404 in the OSC objects having photometry only in two filters are
objects in BRI). more numerous than those having photometry in three or four
We downloaded the data from the GitHub page9 of the Astrocats filters. Therefore, the less filters we use, the larger sample of SNe
project on 2018 June. The complete data set of 45 162 objects is candidates we obtain. First, we tried to use only two Bessel’s filters.
located at http://snad.space/osc/sne.tar.lzma. Prior to applying the filter transformation, we approximated the LCs
with Gaussian processes (see Section 3.2). To evaluate the quality
of transformation, with two filters only, we chose a few objects with
3 P R E - P RO C E S S I N G
LCs available in both, Sloan and Bessel’s filters, and compared
In this section, we describe how to get features for ML from the
OSC LCs. The pre-processing procedure includes several steps that
10 http://www.sdss3.org/dr8/algorithms/sdssUBVRITransform.php
11 http://www.cadc-ccda.hia-iha.nrc-cnrc.gc.ca/en/community/STETSON/
9 https://github.com/astrocatalogs/ index.html

MNRAS 489, 3591–3608 (2019)

3594 M. V. Pruzhinskaya et al.

Downloaded from https://academic.oup.com/mnras/article/489/3/3591/5554761 by Universidad Nacional Autonoma de Mexico user on 04 June 2024
Figure 2. Light curves of SN2005M. Crosses are the observations in gri filters (Contreras et al. 2010; Ganeshalingam et al. 2010; Silverman et al. 2012).
Solid and dashed lines are the approximated and transformed LCs from the Bessel’s BR and BRI to gri filters, respectively.

the transformed gri with the original ones. As can be seen from unrealistic oscillations. To prevent it we set a lower limit on li as the
the Fig. 2, the results of comparison are unsatisfactory for i filter. maximum time interval between two neighbouring observations,
This indicated that at least one more filter had to be added in the but not larger than 60 d. Also, MULTIVARIATE GAUSSIAN PROCESS
analysis. The same test showed that three filters (BRI) are enough kernel includes six constants, three of which are unit variances of
to adequately reproduce gri LCs (Fig. 2). Since with three filters basis processes and three others describe their pairwise correlations.
the equations become over-determined, we used the least-square Totally, MULTIVARIATE GAUSSIAN PROCESS has nine parameters to
method to solve equation (1). be fitted.
Despite the fact that the transformation between the filters Prior to applying the GP approximation, we prepare the
depends on the spectrum of an object, and Lupton’s equations are data (Fig. 1). First, we transform the magnitudes given by the OSC
derived for stars, not for SNe, the Fig. 2 shows quite good agreement to fluxes and perform all further analysis in the flux space only. Since
between transformed and original LCs. measurements remote in time from the maximum (mainly the upper
limits or host detection) could potentially affect the GP behaviour,
including the main part of LC around maximum, for each object we
3.2 Light curve approximation take only the points in the interval [ −240, +240] d relative to the

Traditionally, ML algorithms require a homogeneous input data maximum in r, r , R filter depending on the sub-sample. The Julian
matrix which, unfortunately, is not the case with SNe. A commonly dates are rounded to integers. We also implement 1-d time-binning
used technique to transform unevenly distributed data into a uniform to the data.
grid is to approximate them with Gaussian processes (Rasmussen & In every bin, the flux y and its error σ are derived from n
Williams 2005). Usually, each LC is approximated by GP indepen- observations {yi , σ i } as follows (Agekian 1972):
dently. However, in this study we use a MULTIVARIATE GAUSSIAN 1
PROCESS12 approximation. For each object it takes into account the wi ≡ ,
σi2
correlation between LCs in different bands, approximating the data 
by GP in all filters in a one global fit (for details see Kornilov w≡ wi ,
et al., in preparation). With this technique, we can reconstruct the wi yi
missing parts of LC from its behaviour in other filters. For example, y= ,
w
in Fig. 11 maximum in g filter is reproduced from the r, i LCs.
This correlation does not rely on any physical assumptions about wi (yi − y)2
σ  ≡ ,
LC shape. As an approximation range we chose [ −20, +100] d. w (n − 1)
We also extrapolated the GP approximation to fill this range if
σw ≡ w−1/2 ,
needed. Once the GP approximation becomes negative, it is zeroed
σ , σ  > σw ,
till infinity. σ = 1 (2)
Gaussian process is based on the so-called kernel, a func- 2
(σ  + σ w ) , σ  ≤ σw ,
tion describing the covariance between two observations. The where w i is the weight of observation, w is the sum of weights,
kernel used in our implementation of MULTIVARIATE GAUSSIAN σ  is the mean error, σ w is the error of the weighted mean. If
PROCESS
 is composed
 of three radial-basis functions ki (t1 , t2 ) = the mean error is larger than the error of the weighted mean, then
exp − (t22−tl 21 ) , where i denotes the photometric band, and li are
2
observation errors are probably underestimated or the object is very
i
the parameters of Gaussian process to be found from the LC variable during the considered time interval. Upper limits are taken
approximation. These length parameters describe the characteristic into account only if there are no detections in the bin. In this case,
time-scale of correlation between observations. If the value of li we keep the most conservative upper limit, i.e. the one with the
is too small the approximated LC will be overfitted and can show smallest flux.
Since for each object, the OSC assembles the photometry ob-
tained by different telescopes with different limited magnitudes, a
12 https://github.com/matwey/gp-multistate-kernel lot of upper limits appeared in between or even simultaneously

MNRAS 489, 3591–3608 (2019)

 SNAD: SuperNova Anomaly Detection 3595

Downloaded from https://academic.oup.com/mnras/article/489/3/3591/5554761 by Universidad Nacional Autonoma de Mexico user on 04 June 2024
Figure 3. An example of multicolour LC explaining the pre-processing procedure. The crosses with error bars denote the real photometric detections. Coloured
triangles are the original upper limits (yup ) which are either later than the latest real detection or earlier than the earliest real detection. They are transformed
into the observations with y = 0 ± 3yup (crosses with the thin error bars). Black triangles are upper limits which are ignored. GP approximation of the crosses
is shown by solid lines.

Figure 4. (a) Fraction of objects from our sample with at least one spectrum in the OSC; (b) Distribution of these objects by the OSC types.

with the real detections. This could also have an undesirable interval relative to the maximum. These values were used as features
impact on the Gaussian processes approximation. Therefore, for for the ML algorithm (Section 4).
each filter we keep only those upper limits which are later than Our final sample consists of 1999 objects, 30 per cent of which
the latest real detection or earlier than the earliest real detection. have at least one spectrum in the OSC (see Fig. 4). The distribution
Furthermore, we reassign the values of these upper limits yup : of these objects by astrophysical types is also shown in Fig. 4. The
the new values are zeros with error equal to 3 × yup . This is classification is extracted from the OSC without any verification,
done to decrease the influence of too high upper limits on the GP it can be photometric or based on one spectrum only. Less than
approximation and to force it to vanish for very early and very late 5 per cent of our sample have <20 photometric points in all three
times. Some particular aspects of pre-processing are illustrated in filters. The distributions of objects by redshift and by number of
Fig. 3. photometric points for the three sub-samples are shown in Figs 5
Once the MULTIVARIATE GAUSSIAN PROCESS approximation and 6. The Fig. 5 contains only 1624 objects which significantly
was done, we visually inspected the resulting LCs. Those SNe exceeds the number of objects with the OSC spectra. The reason for
with unsatisfactory approximation were removed from the sample such discrepancy is that, first, the OSC collects also the photometric
(mainly the objects with bad photometric quality). The remaining redshifts and, secondly, not all spectroscopically confirmed SNe
BRI approximated LCs were then transformed to gri (Section 3.1). have the spectrum available in the public domain (for details, see
We consider the LCs in the observer frame. Since each object has Guillochon et al. 2017).
its own flux scale due to the different origin and different distance,
we normalized the flux vector by its maximum value. Based on
3.3 Dimensionality reduction
the results of this approximation, for each object we extracted the
kernel parameters, the log-likelihood of the fit, LC maximum and After the approximation procedure, each object has 374 features:
normalized photometry in the range of [−20, +100] d with 1-d 121 × 3 normalized fluxes, the LC flux maximum, nine fitted

MNRAS 489, 3591–3608 (2019)

3596 M. V. Pruzhinskaya et al.

Downloaded from https://academic.oup.com/mnras/article/489/3/3591/5554761 by Universidad Nacional Autonoma de Mexico user on 04 June 2024
Figure 7. Isolation forest structure. Forest consists of the independent
decision trees. To build a branching in a tree a random feature and a random
splitting are selected. The tree is built until each object of a sample is isolated
in a separate leaf – the shorter path corresponds to a higher anomaly score
which is also illustrated by the colour. For each object, the measure of its
normality is a function of the depths of the leaves into which it is isolated.

Figure 5. Distribution of objects from our sample by the redshift for three
sub-samples and in total. The redshift is available for 1624 objects only. partitioning tree similar to the widely known Kd-tree (Bentley
1975). However, in contrast to the Kd-tree, a space coordinate (a
feature) and a split value are selected at random for every node
of the isolation tree. This algorithm leads to an unbalanced tree
unsuitable for efficient spatial search. However, the tree has the
following important property: a path distance between the root and
the leaf is shorter on average for points distant from ‘normal’ data.
This allows us to construct enough random trees to estimate average
root-leaf path distance for every data sample that we have, and then
rank the data samples based on the path length. The anomaly score,
defined in a range [0, 1], is assigned to each object (see equation 2
in Liu et al. 2008). Then, objects with the highest anomaly score –
outliers – are selected according to the contamination level which
is a hyperparameter of the algorithm. The isolation forest algorithm
is illustrated in Fig. 7.
We run the isolation forest algorithm on 10 data sets obtained
using the same photometric data (Fig. 1):
(A) data set of 364 photometric characteristics (121 × 3 normal-
Figure 6. Distribution of objects from our sample by the number of ized fluxes, the LC flux maximum),
photometric points for three sub-samples and in total.
(B) data set of 10 parameters of the Gaussian process (nine fitted
parameters of the kernel, the log-likelihood of the fit),
(C) eight data sets obtained by reducing 374 features to 2–9 t-
parameters of the Gaussian process kernel, and the log-likelihood
SNE dimensions (Section 3.3).
of the fit.
We apply the outlier detection algorithm not only to the full data For each data set, we obtained a list of outliers. Contamination
set but also to the dimensionality-reduced data. The reason for this levels were set to 1 per cent (20 objects with highest anomaly score)
is that the initial high dimensional feature space can be too sparse for data sets A and B. For all data sets in case C we considered
for the successful performance of the isolation forest algorithm. 2 per cent contamination (40 objects with highest anomaly score).
We applied t-SNE (Maaten & Hinton 2008), a variation of the This larger contamination was chosen to take into account the
stochastic neighbour embedding method (Hinton & Roweis 2003), influence of the dimensionality reduction step in the final data
for the dimensionality reduction of the data. In the t-SNE technique, configuration. Given different representations of the data and the
a non-linear dimensionality reduction mapping is obtained so as stochastic nature of the isolation forest algorithm, the same object
to keep distribution of distances between points undisturbed. This can be assigned a different anomaly score depending on how many
ensures that if a point is anomalous in the sense that it is distant t-SNE dimensions are used. Thus, only those objects which were
from other points in the original data, it remains anomalous in listed within the 2 per cent contamination in at least two of the data
the lower dimension space. As a result of the dimensionality sets in case C are included in Table A1 and subjected to further
reduction, we obtain eight separate reduced data sets corresponding astrophysical analysis. The distribution of objects in each of 10 data
to 2–9 t-SNE features (dimensions). Since t-SNE is a stochastic sets by anomaly score is presented in Fig. 8.
technique we have also taken additional precautions to ensure that An example of the isolation forest algorithm applied to the three-
the resulting outlier list does not depend on the t-SNE initial random dimensional reduced data set is shown in Fig. 9.
state.
5 R E S U LT S
4 I S O L AT I O N F O R E S T
Applying the unsupervised learning to the photometric data ex-
Isolation forest (Liu, Ting & Zhou 2008; Liu et al. 2012) is an tracted from the OSC we found 100 outliers among a total of 1999
ensemble of random isolation trees. Each isolation tree is a space objects (Fig. 1). However, not all of them are necessary anomalies.

MNRAS 489, 3591–3608 (2019)

 SNAD: SuperNova Anomaly Detection 3597
5.1 Peculiar SNe Ia
Type Ia SN is an explosion of a carbon-oxygen white dwarf that
exceeds the Chandrasekhar limit either by matter accretion from a
companion star or by merging with another white dwarf (Whelan &
Iben 1973; Iben & Tutukov 1984; Webbink 1984). SNe Ia are used as
universal distance ladder since their luminosity at maximum light is
approximately the same (Riess et al. 1998; Perlmutter et al. 1999).

Downloaded from https://academic.oup.com/mnras/article/489/3/3591/5554761 by Universidad Nacional Autonoma de Mexico user on 04 June 2024
However, the class of SNe Ia is not homogeneous, for example,
91T-like SNe are on average 0.2–0.3 mag more luminous than
normal SNe Ia, have broader LCs, and different early spectrum
evolution (Filippenko et al. 1992b; Blondin et al. 2012); 91bg-
like SNe are subluminous and fast-declining (Filippenko et al.
1992a); peculiar SNe Iax are spectroscopically similar to SNe Ia,
but have lower maximum light velocities and typically lower peak
magnitudes (Foley et al. 2013). The presence of non-classical SNe Ia
Figure 8. Distribution of objects by anomaly score in 10 data sets described
in Section 4, C 2–C 9 denote C data sets with 2–9 t-SNE dimensions. In each
in cosmological samples may introduce a systematic bias and affect
data set, objects are ordered by score. Black solid and dashed lines denote the cosmological analysis (e.g. Scalzo et al. 2012).
1 per cent and 2 per cent contamination level of outliers, respectively.

5.1.1 SN2002bj
SN2002bj was discovered in NGC 1821 on unfiltered CCD frames
taken with the Puckett Observatory 0.60-m automated patrol tele-
scope on 2002 February 28.06 and March 1.05 UT, and on unfiltered
CCD LOTOSS images taken with the 0.8-m Katzman Automatic
Imaging Telescope on February 28.2 and March 1.2 UT (Puckett
et al. 2002). This SN was a first representative of rapidly evolving
events (Fig. 10). Its LC has a rise time of <7 d followed by a
decline of 0.25 mag d−1 in B band and reaches a peak intrinsic
brightness greater than −18 mag (Poznanski et al. 2010). The
spectra are similar to that of a SN Ia but show the presence of helium
Figure 9. Three-dimensional t-SNE reduced data after application of the and carbon lines. The analysis of archive data after the discovery
isolation forest algorithm. Each point represents an SN LC from the data set
of this object and the subsequent observations revealed other
projected into the three-dimensional space with the coordinates (x1 , x2 , x3 ).
The intensity of the colour indicates the anomaly score for each object as
bright, fast-evolving SNe, e.g. SN1885A, SN1939B, SN2010X,
estimated by the isolation forest algorithm. A darker colour corresponds to SN2015U (Kasliwal et al. 2010; Perets et al. 2011; Shivvers et al.
the objects with higher anomaly scores. 2016). These objects can be produced by the detonation of a helium
shell on a white dwarf, ejecting a small envelope of material
(Poznanski et al. 2010).
That is why, we also subject the outliers to the careful astrophysical
analysis. Using publicly available sources, we collected information
about each outlier and determined to which kind of astrophysical
objects it belongs – given the information we could gather. Among 5.1.2 SN2013cv
the detected outliers there are few known cases of misclassifications, SN2013cv was independently discovered by Zhou et al. (2013) and
representatives of rare classes of SNe (e.g. superluminous SNe, 91T- iPTF (Law et al. 2009) on 2013 May 1.44 UT, see Fig. 11. This
like SNe Ia) and highly reddened objects. We also found that 16 peculiar SN has large peak optical and UV luminosity and show
anomalies classified as SNe by Sako et al. (2018), are likely to be an absence of iron absorption lines in the early spectra. Cao et al.
quasars or stars. (2016b) suggests that SN2013cv is an intermediate case between
LCs with GP approximation for all 1999 objects can be found the normal and super-Chandrasekhar events.
at http://snad.space/osc/ and those who considered anomalous
according to the criteria described in the previous section are
listed in Table A1. Names and equatorial coordinates of outliers
5.1.3 SN2016bln
are shown in Columns 1–3; types in Column 4. CMB redshifts
are presented in Column 5. Columns 6–8 contain the names and SN2016bln/iPTF16abc discovered by the iPTF on 2016 April l3.36
equatorial coordinates of the corresponding host galaxies. Host UT (Cenko et al. 2016; Miller et al. 2016) and classified by our
morphological types are displayed in Column 9. Columns 10 and code as outlier, belongs to the 91T-like SNe Ia subtype (see Fig. 12).
11 contain the separation between centre of the host and object The transitional and nebular spectrum of SN2016bln appear similar
in angular seconds and kiloparsecs, respectively (to calculate the to the normal SN2011fe as well as to overluminous SNe 1991T
angular diameter distance we use a flat CDM cosmology with and 1999aa (Dhawan et al. 2018). Early-time observations show a
H0 = 70 km s−1 Mpc−1 ,  = 0.7). We give our comments and peculiar rise time, non-evolving blue colour, and unusual strong C II
short description of each object in Column 12. References are in absorption. These features can be explained by the ejecta interaction
Column 13. The most interesting of these objects are described with nearby, unbound material or/and significant 56 Ni mixing within
below. the SN ejecta (Miller et al. 2018).

MNRAS 489, 3591–3608 (2019)

3598 M. V. Pruzhinskaya et al.

Downloaded from https://academic.oup.com/mnras/article/489/3/3591/5554761 by Universidad Nacional Autonoma de Mexico user on 04 June 2024
Figure 10. LCs in gri filters of peculiar SN Ia 2002bj (Poznanski et al. 2010). Solid lines are the results of our approximation by MULTIVARIATE GAUSSIAN
PROCESS. The LCs in gri filters are obtained from the Bessel’s BRI by filter transformation (Section 3.1), thus the observations are absent on the plot. The
vertical line denotes the moment of maximum in R filter.

Figure 11. LCs in gri filters of peculiar SN Ia 2013cv (Yaron & Gal-Yam 2012; Cao et al. 2016b). Solid lines are the results of our approximation by
MULTIVARIATE GAUSSIAN PROCESS. The vertical line denotes the moment of maximum in r filter.

Figure 12. LCs in gri filters of 91T-like SN Ia 2016bln (Miller et al. 2018). Solid lines are the results of our approximation by MULTIVARIATE GAUSSIAN
PROCESS. The vertical line denotes the moment of maximum in r filter.

5.2 Peculiar SNe II hydrogen lines are the most prominent in the spectra of SNe II.
Based on the shape of LCs Type II SNe have historically been
Type II SNe arise from the core collapse of massive stars at the final
divided into the Type IIL (linear) and Type IIP (plateau) subtypes,
stage of their evolution. The radius of these stars can be several
however the following studies revealed a continuity in LC slopes of
hundred times greater than the solar radius, and their extremely
Type II SNe (Anderson et al. 2014; Sanders et al. 2015a).
tenuous envelopes contain large amounts of hydrogen. That is why,

MNRAS 489, 3591–3608 (2019)

 SNAD: SuperNova Anomaly Detection 3599

Downloaded from https://academic.oup.com/mnras/article/489/3/3591/5554761 by Universidad Nacional Autonoma de Mexico user on 04 June 2024
Figure 13. LCs in gri filters of peculiar SN II 2013ej (Yuan et al. 2016). Solid lines are the results of our approximation by MULTIVARIATE GAUSSIAN PROCESS.
The vertical line denotes the moment of maximum in r filter.

Figure 14. LCs in gri filters of peculiar SN II 2016ija (Tartaglia et al. 2018). Solid lines are the results of our approximation by MULTIVARIATE GAUSSIAN
PROCESS. The vertical line denotes the moment of maximum in r filter.

5.2.1 SN2013ej leading to a classification as a highly extinguished Type II SN.

The colour excess from the host galaxy NGC 1532 is E(B −
LC of SN2013ej, discovered by the Lick Observatory Supernova
V)host = 1.95 ± 0.15 mag (Tartaglia et al. 2018). Moreover,
Search on 2013 July 25.45 UT (Kim et al. 2013), appears inter-
SN2016ija is brighter than usual SNe II (see fig. 6 of Tartaglia et al.
mediate between those of Type IIP and IIL SNe (see Fig. 13). The
2018).
event has a higher peak luminosity, a faster post-peak decline, and a
shorter plateau phase compared to the normal Type IIP SN 1999em.
The radioactive 56 Ni mass is 0.02 M , which is significantly lower 5.3 Superluminous SNe
than for typical SNe IIP (Huang et al. 2015). The source exhibits
Superluminous SNe are SNe with an absolute peak magnitude M <
signs of substantial geometric asphericity, X-rays from persistent
−21 mag in any band. According to Gal-Yam (2012) SLSN can be
interaction with circumstellar material (CSM), thermal emission
divided into three broad classes: SLSN-I without hydrogen in their
from warm dust (Mauerhan et al. 2017).
spectra, hydrogen-rich SLSN-II that often show signs of interaction
with CSM, and finally, SLSN-R, a rare class of hydrogen-poor
events with slowly evolving LCs, powered by the radioactive decay
5.2.2 SN2016ija of 56 Ni. SLSN-R are suspected to be pair-instability SNe: the deaths
This SN was discovered on 2016 November 21.19 UT (Tartaglia, of stars with initial masses between 140 and 260 solar masses.
Sand & Valenti 2016; see Fig. 14) during 1-d cadence SN search In our outlier list in Table A1 there are four SLSN: SDSS-II SN
for very young transients in the nearby Universe (DLT40). Using 17789, SN2015bn, PTF10aagc, SN2213-1745.
SNID (Blondin & Tonry 2007), it was first suggested to be an
early time 91T-like SN Ia with few features and red continuum.
5.3.1 SN2213-1745
It has been also associated to the outburst in an obscured lumi-
nous blue variable, an intermediate luminosity red transient or a SN2213-1745 was discovered at z = 2.046 by the Canada–France–
luminous red nova (Blagorodnova et al. 2016). The subsequent Hawaii Telescope Legacy Survey (Fig. 15). It belongs to the SLSN-
spectroscopic follow-up revealed broad Hα and calcium features, R events. Cooke et al. (2012) suggested that SN 2213-1745 may be

MNRAS 489, 3591–3608 (2019)

3600 M. V. Pruzhinskaya et al.

Downloaded from https://academic.oup.com/mnras/article/489/3/3591/5554761 by Universidad Nacional Autonoma de Mexico user on 04 June 2024
  
Figure 15. LCs in g r i filters of superluminous SN SN2213-1745 (Cooke et al. 2012). Solid lines are the results of our approximation by MULTIVARIATE

GAUSSIAN PROCESS. The vertical line denotes the moment of maximum in r filter.

Figure 16. LCs in gri filters of superluminous SN PTF10aagc (De Cia et al. 2018). Solid lines are the results of our approximation by MULTIVARIATE GAUSSIAN
PROCESS. The vertical line denotes the moment of maximum in r filter.

  
Figure 17. LCs in g r i filters of AGN SN2006kg (Sako et al. 2018). Solid lines are the results of our approximation by MULTIVARIATE GAUSSIAN PROCESS.

The vertical line denotes the moment of maximum in r filter.

powered by the radiative decay of a 4–7 M of synthesized 56 Ni, the latter spectra revealed a broad Hα and the corresponding
and implied a progenitor with an estimated initial mass of 250 M . weak, but detected Hβ (Yan et al. 2015). This particularity makes
PTF10aagc clearly distinct from others SLSN-I. Such spectral
behaviour can be explained by interaction between SLSN-I ejecta
5.3.2 PTF10aagc
and a H-rich circumstellar material at late times (Yan et al. 2015).
The high peak luminosity (Lbol, peak = 1043.7 erg s−1 ) and the The host of PTF10aagc is bright and shows clear morphological
absence of hydrogen lines in early spectrum allowed to attribute structure suggesting a possible ongoing merger (Perley et al.
PTF10aagc to SLSN-I (De Cia et al. 2018; see Fig. 16). However, 2016).

MNRAS 489, 3591–3608 (2019)

 SNAD: SuperNova Anomaly Detection 3601

Downloaded from https://academic.oup.com/mnras/article/489/3/3591/5554761 by Universidad Nacional Autonoma de Mexico user on 04 June 2024
Figure 18. LCs in gri filters of binary microlensing event Gaia16aye (http://gsaweb.ast.cam.ac.uk/alerts/alert/Gaia16aye/followup). Solid lines are the results
of our approximation by MULTIVARIATE GAUSSIAN PROCESS. The vertical line denotes the moment of maximum in r filter.

5.4 Misclassified objects 6 CONCLUSIONS

5.4.1 SN2006kg The development of large sky surveys has led to a discovery of a
huge number of SNe and SN candidates. Among the SNe discovered
SN2006kg was first classified as a possible Type II SN (Bassett et al. every year, only 10 per cent have spectroscopic confirmation. The
2006; see Fig. 17). It is also appeared as Type II spectroscopically amount of astronomical data increases dramatically with time and
confirmed SN in table 6 of Sako et al. (2008). However, further is already beyond human capabilities. The astronomical community
analysis of 3.6-m New Technology Telescope spectrum revealed already has dozens of thousands of SN candidates, and LSST
that SN2006kg is an AGN (Östman et al. 2011; Sako et al. 2018). survey (LSST Science Collaboration et al. 2009) will discover over
It is interesting that SN2006kg continues to appear as SN in host 10 million SNe in the forthcoming decade. Only a small fraction
studies (Hakobyan et al. 2012) and was even in a set of 12 well- of them will receive a spectroscopic confirmation. This motivates a
observed events that were used as Type II SN templates (Okumura considerable effort in photometric classification of SNe by types
et al. 2014). The object is not in the WISE AGN Catalogue (Assef using machine learning algorithms. There is, however, another
et al. 2018) that consists of >20 millions AGN candidates. aspect of the problem: any large photometric SN data base would
suffer from the non-SN contamination (novae, kilonovae, GRB
afterglows, AGNs, etc.). Moreover, the data base will inevitably
5.4.2 Gaia16aye contain the astronomical objects with unusual physical properties –
Gaia16aye (Bakis et al. 2016) is an object with the most non-SN- anomalies. Finding such objects and studying them in detail is very
like behaviour among our set of outliers (Fig. 18). In Wyrzykowski important and constitutes the main goal of this paper.
et al. (2016), it was reported that Gaia16aye is a binary microlensing The analysis presented here is based on the photometric data
event – gravitational microlensing of binary systems – the first ever extracted from the OSC (Guillochon et al. 2017). The use of real
discovered towards the Galactic Plane. data allows us to reveal a lot of caveats in observations at the pre-
processing stage – many of which are not normally present in the
simulated data. After pre-processing, we obtain 1999 SNe with LCs
  

5.4.3 Possible misclassified objects either in gri or in g r i or in BRI filters approximated by Gaussian
processes. We consider 10 different data sets: one that includes
Our analysis also reveals that 16 objects classified as pSN by Sako the approximated photometric observations (A), another with the
et al. 2018, where a prefix ‘p’ indicates a purely photometric type, parameters of Gaussian process only (B) and eight data sets were the
are likely to be stars or quasars. First, we do not find any signature information in GP photometry and GP parameters were summarized
of SNe on the corresponding multicolour LCs. Then, according to via dimensionality reduction using t-SNE (dimension varying from
SDSS DR1513 type of SDSS-II SN 5314, SDSS-II SN 14170, SDSS- 2–9, case C).
II SN 15565, SDSS-II SN 13725, SDSS-II SN 13741, SDSS-II SN We apply the isolation forest algorithm to all data sets, con-
19699, SDSS-II SN 18266, SDSS-II SN 4226, SDSS-II SN 2809, sidering a 1 per cent contamination for cases A/B and 2 per cent
SDSS-II SN 6992 is denoted as star. Moreover, all these objects can contamination for all data sets in case C. We visually checked all
be found in Pan-STARRS Catalogue with Pan-STARRS magnitudes objects identified in cases A and B. We also checked all the objects
equal or even brighter than those on the corresponding LCs. which were identified as anomalous in at least two of the data sets
The other objects SDSS-II SN 1706, SDSS-II SN 17756, SDSS- in case C. As a result, we find 100 outliers, 40 from cases A/B and
II SN 17339, SDSS-II SN 17509, SDSS-II SN 4652, SDSS-II SN 60 which were identified in at least two data sets of case C. Among
19395 have a BOSS (Smee et al. 2013) spectrum with class ‘QSO’ these, 19 objects were identified by both strategies, with and without
and have high redshifts (see Table A1). dimensionality reduction. Our final validation analysis resulting in
81 objects which were carefully studied with the use of publicly
available information. Among these, there are four superluminous
13 http://skyserver.sdss.org/dr15/en/tools/explore/summary.aspx SNe (SDSS-II SN 17789, SN2015bn, PTF10aagc, SN2213-1745),

MNRAS 489, 3591–3608 (2019)

3602 M. V. Pruzhinskaya et al.
non-classical Type Ia SNe (91T-like SNe 2016bln, PS15cfn, SNLS- grant 18-32-00553 for preparing the OSC data. E. E. O. Ishida
03D1cm; peculiar SN2002bj and SN2013cv), two unusual Type II acknowledges support from CNRS 2017 MOMENTUM grant
SNe which anomalous multicolour LC behaviour can be due to and Foundation for the advancement of theoretical physics and
the environment (SN2013ej, SN2016ija), one AGN – SN2006kg, Mathematics ‘BASIS’. A. Volnova acknowledges support from RSF
and one binary microlensing event Gaia16aye. We also find that grant 18-12-00522 for analysis of interpolated LCs. We used the
16 anomalies classified as SNe by Sako et al. (2018), which are equipment funded by the Lomonosov Moscow State University Pro-
likely to be stars (SDSS-II SN 5314, SDSS-II SN 14170, SDSS- gram of Development. The authors acknowledge the support from

Downloaded from https://academic.oup.com/mnras/article/489/3/3591/5554761 by Universidad Nacional Autonoma de Mexico user on 04 June 2024
II SN 15565, SDSS-II SN 13725, SDSS-II SN 13741, SDSS-II SN the Program of Development of M.V. Lomonosov Moscow State
19699, SDSS-II SN 18266, SDSS-II SN 4226, SDSS-II SN 2809, University (Leading Scientific School ‘Physics of stars, relativistic
SDSS-II SN 6992) or quasars (SDSS-II SN 1706, SDSS-II SN objects and galaxies’). This research has made use of NASA’s
17756, SDSS-II SN 17339, SDSS-II SN 17509, SDSS-II SN 4652, Astrophysics Data System Bibliographic Services and following
SDSS-II SN 19395). However, without careful spectral analysis it PYTHON software packages: NUMPY (van der Walt, Colbert &
is difficult to distinguish a high-redshift SN against a background Varoquaux 2011), MATPLOTLIB (Hunter 2007), SCIPY (Jones et al.
galaxy or from quasar activity. As a confirmation of the robustness 2001), PANDAS (McKinney 2010), and SCIKIT-LEARN (Pedregosa
of the pipeline used here, we note that of the nine objects identified et al. 2011).
as outliers in all data sets of case C, five are misclassifications and
one is an extreme case of bad photometry.
In summary, the isolation forest analysis identified 81 potentially REFERENCES
interesting objects, from which 27 (33 per cent) where confirmed Agekian T., 1972, Fundamentals of the theory of errors for astronomers and
to be non-SN events or representatives of the rare SN classes. physicists. Nauka, Moscow
Found anomalies correspond to 1.4 per cent of the original data Aldering G. et al., 2002, Proc. Conf. Ser. SPIE, Bellingham. p. 61
set of approximately 2000 objects which was identified demanding Anderson J. P. et al., 2014, ApJ, 786, 67
significantly less resources than a manual search would entail. Assef R. J., Stern D., Noirot G., Jun H. D., Cutri R. M., Eisenhardt P. R. M.,
Among these objects, we report for the first time the 16 star/quasar- 2018, ApJS, 234, 23
like objects misclassified as SNe. Astier P. et al., 2006, A&A, 447, 31
Bakis V. et al., 2016, Astron. Telegram, 9376
It is important to note that these results are not expected to
Ball N. M., Brunner R. J., 2010, Int. J. Mod. Phys. D, 19, 1049
be complete. For example, there are known SLSN which were
Barbon R., Buondı́ V., Cappellaro E., Turatto M., 1999, A&AS, 139, 531
not identified as outliers in our search, as well as one object Baron D., Poznanski D., 2017, MNRAS, 465, 4530
(SN1000 + 0216) at very high redshift which was identified as Bassett B. et al., 2006, Cent. Bur. Electron. Telegrams, 688, 1
anomalous in only one of the data sets for case C – and consequently Bassett B. et al., 2007, Cent. Bur. Electron. Telegrams, 1079, 1
was not included in our final list. This is a natural consequence of Bazin G. et al., 2011, A&A, 534, A43
the pre-processing analysis we chose to adopt, where the objects Beck R., Lin C. A., Ishida E. E. O., Gieseke F., de Souza R. S., Costa-Duarte
are mainly characterized by their LC shape (all photometric features M. V., Hattab M. W., Krone-Martins A., 2017, MNRAS, 468, 4323
were normalised). In this context, differences in intrinsic brightness Bellm E. C. et al., 2019, PASP, 131, 018002
only marginally affect our final results. Another source of false Ben-Ami S. et al., 2014, ApJ, 785, 37
Bentley J. L., 1975, Commun. ACM, 18, 509
negatives can be traced back to GP approximations, the typical
Bessell M. S., 1990, PASP, 102, 1181
example being Gaia16aye (Fig. 18). From a visual inspection of
Betoule M. et al., 2014, A&A, 568, A22
its observed photometric points this object is obviously not an SN. Blagorodnova N., Neill J. D., Kasliwal M., Walters R., Adams S. M., 2016,
However, it appeared only in three of the eight possible data sets Astron. Telegram, 9787
in case C. A more detailed analysis of its GP approximation (solid Blondin S., Tonry J. L., 2007, ApJ, 666, 1024
line in Fig. 18), reveals that the information input to the ML model Blondin S. et al., 2012, AJ, 143, 126
was much smoother than one would expect. As a consequence, the Bloom J. S. et al., 2012, PASP, 124, 1175
algorithm struggles to separate it from other slow declining events. Bose S. et al., 2015a, MNRAS, 450, 2373
Nevertheless, the above results provide clear evidence of the Bose S. et al., 2015b, ApJ, 806, 160
effectiveness of automated anomaly detection algorithms for pho- Branch D. et al., 2006, PASP, 118, 560
Brunel A., Pasquet J., Pasquet J., Rodriguez N., Comby F., Fouchez D.,
tometric SN LC analysis. In this work, we used data from the OSC
Chaumont M., 2019, preprint (arXiv:1901.00461)
in order to provide a proof of concept. Although this is not a big data
Campbell H., Blagorodnova N., Fraser M., Gilmore G., Hodgkin S.,
sample, it does allow us to search for independent information in the Koposov S., Walton N., Wyrzykowski L., 2014, in Wozniak P. R.,
literature on which we could confirm our findings. This approach Graham M. J., Mahabal A. A., Seaman R., eds, The Third Hot-wiring
to the analysis of photometric LCs will be paramount for future the Transient Universe Workshop. Santa Fe, New Mexico, p. 43
astronomical surveys like LSST, which will not be able to afford Cao Y., Nugent P. E., Kasliwal M. M., 2016a, PASP, 128, 114502
a manual research or the possibility to overlook interesting objects Cao Y. et al., 2016b, ApJ, 823, 147
deviating from the bulk of the data – where the most interesting Cavuoti S. et al., 2015, MNRAS, 452, 3100
physics resides. Cenko S. B., Cao Y., Kasliwal M., Miller A. A., Fremling C., West M.,
The code of this work and the data are available at http://snad.s Gregg M., Kulkarni S. R., 2016, Astron. Telegram, 8909
Chambers K. C. et al., 2016, preprint (arXiv:1612.05560)
pace/osc/.
Charnock T., Moss A., 2017, ApJ, 837, L28
Chen X., Ishwaran H., 2012, Genomics, 99, 323
AC K N OW L E D G E M E N T S Chiaki G., Yoshida N., Kitayama T., 2013, ApJ, 762, 50
Cikota A. et al., 2016, Astron. Telegram, 9889
M. Pruzhinskaya and M. Kornilov are supported by RFBR grant Contreras C. et al., 2010, AJ, 139, 519
according to the research project 18-32-00426 for outlier analysis Cooke J. et al., 2012, Nature, 491, 228
and LCs approximation. K. Malanchev is supported by RBFR Criscia C., Ghattasb B., Pererac G., 2012, Ecological Modelling, 240, 113

MNRAS 489, 3591–3608 (2019)

 SNAD: SuperNova Anomaly Detection 3603
Dark Energy Survey Collaboration et al., 2016, MNRAS, 460, 1270 Mauerhan J. C. et al., 2017, ApJ, 834, 118
De Cia A. et al., 2018, ApJ, 860, 100 McCauliff S. D. et al., 2015, ApJ, 806, 6
Dhawan S. et al., 2018, MNRAS, 480, 1445 McKinney W., 2010, in van der Walt S., Millman J., eds, Proc. 9th Python
Dong S. et al., 2016, Science, 351, 257 in Science Conference. p. 51
du Buisson L., Sivanandam N., Bassett B. A., Smith M., 2015, MNRAS, Miller A. A. et al., 2016, Astron. Telegram, 8907
454, 2026 Miller A. A. et al., 2018, ApJ, 852, 100
Dubost F., Yilmaz P., Adams H., Bortsova G., Ikram M. A., Niessen W., Möller A., de Boissière T., 2019, preprint (arXiv:1901.06384)
Vernooij M., de Bruijne M., 2019, NeuroImage, 185, 534 Möller A. et al., 2016, J. Cosmol. Astropart. Phys., 2016, 008

Downloaded from https://academic.oup.com/mnras/article/489/3/3591/5554761 by Universidad Nacional Autonoma de Mexico user on 04 June 2024
Filippenko A. V. et al., 1992a, AJ, 104, 1543 Monard L. A. G., 2006, IAU Circ., 8666
Filippenko A. V. et al., 1992b, ApJ, 384, L15 Morlino G., 2017, Handbook of Supernovae. Springer-Verlag, Berlin, p.
Folatelli G. et al., 2006, ApJ, 641, 1039 1711
Folatelli G. et al., 2013, ApJ, 773, 53 Muthukrishna D., Parkinson D., Tucker B., 2019, preprint (arXiv:1903.025
Foley R. J. et al., 2013, ApJ, 767, 57 57)
Foley R. J. et al., 2018, MNRAS, 475, 193 Nagakura T., Hosokawa T., Omukai K., 2009, MNRAS, 399, 2183
Fukugita M., Ichikawa T., Gunn J. E., Doi M., Shimasaku K., Schneider D. Narayan G. et al., 2018, ApJS, 236, 9
P., 1996, AJ, 111, 1748 Nicholl M. et al., 2016, ApJ, 828, L18
Gal-Yam A., 2012, Science, 337, 927 Nomoto K., Kobayashi C., Tominaga N., 2013, ARA&A, 51, 457
Ganeshalingam M. et al., 2010, ApJS, 190, 418 Nun I., Pichara K., Protopapas P., Kim D.-W., 2014, ApJ, 793, 23
Ganeshalingam M., Li W., Filippenko A. V., 2013, MNRAS, 433, Okumura J. E. et al., 2014, PASJ, 66, 49
2240 Pasquet J., Pasquet J., Chaumont M., Fouchez D., 2019, A&A, 627,
Goldstein D. A. et al., 2015, AJ, 150, 82 15
Guillochon J., Parrent J., Kelley L. Z., Margutti R., 2017, ApJ, 835, 64 Pastorello A. et al., 2016, MNRAS, 456, 853
Guy J. et al., 2010, A&A, 523, A7 Pearson K. A., Palafox L., Griffith C. A., 2018, MNRAS, 474,
Hakobyan A. A., Adibekyan V. Z., Aramyan L. S., Petrosian A. R., Gomes 478
J. M., Mamon G. A., Kunth D., Turatto M., 2012, A&A, 544, A81 Pedregosa F. et al., 2011, J. Mach. Learn. Res., 12, 2825
Hamuy M., Pinto P. A., 2002, ApJ, 566, L63 Perets H. B., Badenes C., Arcavi I., Simon J. D., Gal-yam A., 2011, ApJ,
Hamuy M. et al., 2006, PASP, 118, 2 730, 89
Heinis S. et al., 2016, ApJ, 821, 86 Perley D. A. et al., 2016, ApJ, 830, 13
Henrion M., Hand D. J., Gandy A., Mortlock D. J., 2013, Statist. Anal. Data Perlmutter S. et al., 1999, ApJ, 517, 565
Min. ASA Data Sci. J., 6, 53 Poznanski D. et al., 2010, Science, 327, 58
Hildebrandt H. et al., 2010, A&A, 523, A31 Pritchet C. J., SNLS Collaboration, 2005, in Wolff S. C., Lauer T. R., eds,
Hinton G. E., Roweis S. T., 2003, in Backer S., Thrun S., Obermayer K., eds, ASP Conf. Ser. Vol. 339, Observing Dark Energy. Astron. Soc. Pac.,
2002 Neural al Information Processing Systems. MIT Press, Cambridge, San Francisco, p. 60
MA, p. 857 Puckett T., Newton J., Papenkova M., Li W. D., 2002, Int. Astron. Union
Holoien T. W.-S. et al., 2019, MNRAS, 484, 1899 Circ., 7839, 1
Hsiao E. Y. et al., 2013, Astron. Telegram, 5678 Quang D., Xie X., 2016, Nucleic Acids Res., 44, e107
Huang F. et al., 2015, ApJ, 807, 59 Rasmussen C. E., Williams C. K. I., 2005, Gaussian Processes for Machine
Hunter J. D., 2007, Comput. Sci. Eng., 9, 90 Learning (Adaptive Computation and Machine Learning). The MIT
Iben Jr. I., Tutukov A. V., 1984, ApJS, 54, 335 Press, Cambridge, MA
Inserra C. et al., 2012, MNRAS, 422, 1122 Rebbapragada U., Protopapas P., Brodley C. E., Alcock C., 2009, Mach.
Ishida E. E. O., de Souza R. S., 2013, MNRAS, 430, 509 Learn., 74, 281
Ishida E. E. O. et al., 2019, MNRAS, 483, 2 Rest A. et al., 2014, ApJ, 795, 44
Jha S., Riess A. G., Kirshner R. P., 2007, ApJ, 659, 122 Revsbech E. A., Trotta R., van Dyk D. A., 2018, MNRAS, 473,
Jones E., Oliphant T., Peterson P. et al., 2001, SciPy: Open source scientific 3969
tools for Python. Available at: http://www.scipy.org/ Richards J. W., Homrighausen D., Freeman P. E., Schafer C. M., Poznanski
Kaiser N. et al., 2010, in Stepp L. M., Gilmozzi R., Hall H. J., eds, Proc. D., 2012, MNRAS, 419, 1121
Conf. Ser. Vol. 7733 Ground-based and Airborne Telescopes III. SPIE, Riess A. G. et al., 1998, AJ, 116, 1009
Bellingham, p. 77330E Rodrı́guez Ó., Clocchiatti A., Hamuy M., 2014, AJ, 148, 107
Kasliwal M. M. et al., 2010, ApJ, 723, L98 Rubin A., Gal-Yam A., 2016, ApJ, 828, 111
Kessler R. et al., 2010, PASP, 122, 1415 Sako M. et al., 2008, AJ, 135, 348
Kim M. et al., 2013, Cen. Bur. Electron. Telegrams, 3606 Sako M. et al., 2018, PASP, 130, 064002
Kirshner R. P., Kwan J., 1974, ApJ, 193, 27 Sanders N. E. et al., 2015a, ApJ, 799, 208
Law N. M. et al., 2009, PASP, 121, 1395 Sanders N. E., Betancourt M., Soderberg A. M., 2015b, ApJ, 800,
Le Guillou L. et al., 2015, Astron. Telegram, 7102 36
Leloudas G. et al., 2016, Nature Astron., 1, 0002 Sasdelli M. et al., 2016, MNRAS, 461, 2044
Leonard D. C. et al., 2002, AJ, 124, 2490 Scalzo R. et al., 2012, ApJ, 757, 12
Libbrecht M. W., Noble W. S., 2015, Nature Rev. Genetics, 16, 321 Scolnic D. M. et al., 2018, ApJ, 859, 101
Lipunov V. et al., 2010, Adv. Astron., 2010, 349171 Shivvers I. et al., 2016, MNRAS, 461, 3057
Liu F. T., Ting K. M., Zhou Z.-H., 2008, in Giannotti F., Gunopulos D., Silverman J. M. et al., 2012, MNRAS, 425, 1789
Turini F., Zaniolo C., Ramakrishnan N., Wu X., eds, 2008 Eighth IEEE Smee S. A. et al., 2013, AJ, 146, 32
Int. Conf. Data Mining. IEEE Computer Society, Los Alamitos, CA, p. Smith M. et al., 2012, ApJ, 755, 61
413 Smith J. A. et al., 2007, in Sterken C., ed., ASP Conf. Ser. Vol. 364, The
Liu F. T., Ting K. M., Zhou Z.-H., 2012, ACM Trans. Knowl. Discovery Future of Photometric, Spectrophotometric and Polarimetric Standard-
Data, 6, 3:1 ization. Astron. Soc. Pac., San Francisco, p. 91
Lochner M., McEwen J. D., Peiris H. V., Lahav O., Winter M. K., 2016, Sooknunan K. et al., 2018, preprint (arXiv:1811.08446)
ApJS, 225, 31 Stritzinger M. D. et al., 2018, A&A, 609, A134
LSST Science Collaboration et al., 2009, preprint (arXiv:0912.0201) Tartaglia L., Sand D., Valenti S., 2016, Astron. Telegram, 9782
Maaten L. v. d., Hinton G., 2008, J. Mach. Learn. Res., 9, 2579 Tartaglia L. et al., 2018, ApJ, 853, 62

MNRAS 489, 3591–3608 (2019)

3604 M. V. Pruzhinskaya et al.
Thompson S. E., Mullally F., Coughlin J., Christiansen J. L., Henze C. E., Wright D. E. et al., 2015, MNRAS, 449, 451
Haas M. R., Burke C. J., 2015, ApJ, 812, 46 Wyrzykowski Ł., Hodgkin S., Blogorodnova N., Koposov S., Burgon
Tsvetkov D. Y., Pavlyuk N. N., Bartunov O. S., 2005, Astronomy Letters, R., 2012, in Tango P., Thullot W., eds, 2nd Gaia Follow-up Net-
30, 729 work for Solar System Objects. Observatoire de Paris, Paris, France,
Tucker D. L. et al., 2006, Astron. Nachr., 327, 821 p. 21
van der Walt S., Colbert S. C., Varoquaux G., 2011, Comput. Sci. Eng., 13, Wyrzykowski L. et al., 2016, Astron. Telegram, 9507
22 Yan L. et al., 2015, ApJ, 814, 108
Venkatraghavan V., Bron E. E., Niessen W. J., Klein S., 2019, NeuroImage, Yaron O., Gal-Yam A., 2012, PASP, 124, 668

Downloaded from https://academic.oup.com/mnras/article/489/3/3591/5554761 by Universidad Nacional Autonoma de Mexico user on 04 June 2024
186, 518 Yuan F. et al., 2016, MNRAS, 461, 2003
Vilalta R., Ishida E. E. O., Beck R., Sutrisno R., de Souza R. S., Mahabal A., Zheng C. et al., 2008, AJ, 135, 1766
2017, in Bonissone P. P., Fogel D., eds, 2017 IEEE Symposium Series Zhou L. et al., 2013, Cent. Bur. Electron. Telegrams, 3543
on Computational Intelligence (SSCI). IEEE, Los Alamitos, CA, p. 1 Östman L. et al., 2011, A&A, 526, A28
Wang X. et al., 2008, ApJ, 675, 626
Wang X. et al., 2009, ApJ, 699, L139
Webbink R. F., 1984, ApJ, 277, 355
Whelan J., Iben Jr. I., 1973, ApJ, 186, 1007 A P P E N D I X : TA B L E W I T H O U T L I E R S

MNRAS 489, 3591–3608 (2019)

 Table A1. List of outliers and their hosts.
Name α δ Typea zCMB Host name Host α Host δ Host Sep. (”)c Sep. Commentsd References
typeb (kpc)c
Outliers found in eight data sets with different dimensionality reduction
SDSS-II SN 01:09:29.89 −01:05:18.3 ?SN II SDSS J010929.89-010518.3 01:09:29.89 −01:05:18.3 pSN II in Sako et al. (2018); SDSS DR15 Sako et al. (2018)
13112e host photoZ (KD-tree method)
0.755 ± 0.234
SDSS-II SN 23:03:39.48 +00:12:49.9 ?SN II SDSS J230339.49+001249.7 23:03:39.49 +00:12:49.8 pSN II in Sako et al. (2018); SDSS DR15 Sako et al. (2018)
13461 host photoZ (KD-tree method)
0.735 ± 0.080
SDSS-II SN 20:21:17.85 +00:41:04.3 ?SN II/?Star SDSS J202117.84+004104.2 20:21:17.84 +00:41:04.2 pSN II in Sako et al. (2018); host classified Sako et al. (2018)
5314e as star by SDSS DR15
SN2016fbof 01:01:35.54 +17:06:04.3 SN Ia 0.030 GALEXASC 01:01:35.77 +17:06:04.1 3.3 1.98 LC in the OSC has a bad quality Foley et al. (2018)
J010135.75+170604.9
SDSS-II SN 01:46:33.15 +00:51:05.7 ?SN II/?Star SDSS J014633.15+005105.6 01:46:33.15 +00:51:05.6 pSN II in Sako et al. (2018); host classified Sako et al. (2018)
14170e as star by SDSS DR15
SDSS-II SN 00:02:58.11 −01:01:27.8 ?SN II/QSO SDSS J000258.10-010127.8 00:02:58.10 −01:01:27.8 pSN II in Sako et al. (2018); according to Sako et al. (2018)
1706e SDSS DR15 host has BOSS spectrum with
z = 1.551, class = QSO broadline
LSQ13dpae 11:01:12.91 −05:50:52.6 SN II 0.023 LCSB S1492O 11:01:12.46 −05:50:46.0 S 9.4 4.37 Spectroscopically confirmed as SN II using Hsiao et al. (2013)
a near-infrared spectrum (range
800–2500 nm)
SDSS-II SN 01:08:10.42 −00:16:36.9 ?SN II/QSO SDSS J010810.43-001636.9 01:08:10.43 −00:16:37.0 pSN II in Sako et al. (2018); host classified Sako et al. (2018)
17756e as star by SDSS DR15, however, it has a
BOSS spectrum with z = 1.997, class =
QSO broadline
SDSS-II SN 01:00:27.12 +00:35:23.6 ?SN II/?Star SDSS J010027.10+003523.5 01:00:27.10 +00:35:23.5 pSN II in Sako et al. (2018); host classified Sako et al. (2018)
15565 as star by SDSS DR15
Outliers found in seven data sets with different dimensionality reduction
SN2005hof 00:59:24.10 +00:00:09.3 SN Ia 0.062 PGC 1154577 00:59:24.10 +00:00:09.4 Sm/Im 0.1 0.12 In JLA cosmological sample (Betoule et al. Betoule et al. (2014)
2014), not in Pantheon (Scolnic et al. 2018)
SN2016ija 04:12:07.62 −32:51:10.9 SN II 0.003 NGC 1532 04:12:04.33 −32:52:27.2 Sbc 86.8 5.39 Highly obscured SN II [E(B − V)host = Tartaglia et al. (2018)
1.95 ± 0.15 mag]
SDSS-II SN 20:54:41.52 −00:13:33.2 Unknown SDSS J205441.53-001333.1 20:54:41.53 −00:13:33.2 Unknown object in Sako et al. (2018) and Sako et al. (2018)
2050e SN II in the OSC with reference to Sako
et al. (2018)
SDSS-II SN 02:31:22.22 +00:23:09.6 ?SN II/QSO SDSS J023122.22+002309.5 02:31:22.22 +00:23:09.5 pSN II in Sako et al. (2018); host classified Sako et al. (2018)
17339e as star by SDSS DR15, however, it has a
BOSS spectrum with z = 1.132, class =
QSO
Outliers found in six data sets with different dimensionality reduction
f e
SN2007jm , 21:55:38.59 −00:10:36.3 SN IIn 0.090 SDSS J215538.80-001034.1 21:55:38.80 −00:10:34.2 3.8 6.36 According to SDSS DR15 host has BOSS Bassett et al. (2007);
spectrum with z = 0.091, class = galaxy Sako et al. (2018)
star forming
SNLS-06D3gx 14:17:03.23 +52:56:10.5 SN Ia 0.761 SDSS J141703.21+525616.1 14:17:03.21 +52:56:16.1 5.6 41.31 In JLA cosmological sample (Betoule et al. Betoule et al. (2014)
2014), not in Pantheon (Scolnic et al. 2018)
SDSS-II SN 21:06:55.01 +00:53:44.9 ?SN II/?Star SDSS J210655.01+005344.8 21:06:55.01 +00:53:44.9 pSN II in Sako et al. (2018); host classified Sako et al. (2018)
13725 as star by SDSS DR15
SN2016ixf 10:39:44.56 +15:02:04.8 SN Ia 0.067 SDSS J103944.53+150204.7 10:39:44.53 +15:02:04.8 0.4 0.56 Cikota et al. (2016); Foley
et al. (2018)
SN2006ob 01:51:48.14 +00:15:47.9 SN Ia 0.059 UGC 1333 01:51:48.51 +00:15:49.8 Sa 5.9 6.69 In JLA (Betoule et al. 2014) and Betoule et al. (2014)
Pantheon (Scolnic et al. 2018)
SNAD: SuperNova Anomaly Detection

cosmological samples

MNRAS 489, 3591–3608 (2019)

3605

Downloaded from https://academic.oup.com/mnras/article/489/3/3591/5554761 by Universidad Nacional Autonoma de Mexico user on 04 June 2024
 Table A1 – continued
3606

Name
pt α δ Typea zCMB Host name Host α Host δ Host Sep. (”)c Sep. Commentsd References
typeb (kpc)c
SDSS-II SN 00:19:18.93 +01:08:14.3 ?SN II/QSO SDSS J001918.93+010814.2 00:19:18.93 +01:08:14.2 pSN II in Sako et al. (2018); host classified Sako et al. (2018)
17509 as star by SDSS DR15, however, it has a
BOSS spectrum with z = 2.031, class =
QSO
Outliers found in five data sets with different dimensionality reduction
SDSS-II SN 4330 01:44:35.82 −00:10:57.4 ?SN II SDSS J014435.82-001057.3 01:44:35.82 −00:10:57.4 pSN II in Sako et al. (2018); SDSS DR15 Sako et al. (2018)
host photoZ (KD-tree method)
0.732 ± 0.076
SN2005ll 22:28:06.87 −01:07:41.4 SN Ia 0.241 SDSS J222806.92-010742.1 22:28:06.92 −01:07:42.1 1.0 3.90 According to SDSS DR15 host has BOSS Sako et al. (2018)

MNRAS 489, 3591–3608 (2019)

spectrum with z = 0.242, class = galaxy
star forming
SDSS-II SN 20:48:00.40 −01:02:49.5 ?SN II/?Star SDSS J204800.39-010249.4 20:48:00.39 −01:02:49.5 pSN II in Sako et al. (2018); host classified Sako et al. (2018)
13741 as star by SDSS DR15
SDSS-II SN 23:31:23.77 +00:37:45.6 ?SN II SDSS J233123.77+003745.4 23:31:23.77 +00:37:45.5 pSN II in Sako et al. (2018); SDSS DR15 Sako et al. (2018)
M. V. Pruzhinskaya et al.

17292e host photoZ (KD-tree method)

0.595 ± 0.163
SN2006kg 01:04:16.98 +00:46:08.9 AGN 0.230 SDSS J010416.98+004608.7 01:04:16.98 +00:46:08.8 Basing on NTT spectrum classified as Sako et al. (2018)
AGNs by Östman et al. (2011); according
to SDSS DR15 host has BOSS spectrum
with z = 0.231, class = galaxy starburst
Outliers found in four data sets with different dimensionality reduction
SDSS-II SN 4652 02:29:49.69 −00:40:11.4 ?SN II/QSO SDSS J022949.69-004011.3 02:29:49.69 −00:40:11.4 pSN II in Sako et al. (2018); according to Sako et al. (2018)
SDSS DR15 host has BOSS spectrum with
z = 0.673, class = QSO
SN2002bj 05:11:46.41 −15:08:10.8 SN Ia pec/SN Ib 0.012 NGC 1821 05:11:46.11 −15:08:04.9 Im 7.3 1.80 Bright, fast-evolving SN with low-mass Poznanski et al. (2010)
pec ejecta, helium and carbon lines in spectra
SDSS-II SN 21:48:02.39 −00:07:07.5 ?SN II SDSS J214802.31-000710.0 21:48:02.31 −00:07:10.1 pSN II in Sako et al. (2018); SDSS DR15 Sako et al. (2018)
13589e host photoZ (KD-tree method)
0.199 ± 0.032
SDSS-II SN 20:04:11.38 −00:32:01.1 ?SN II SDSS J200411.38-003200.9 20:04:11.38 −00:32:01.0 pSN II in Sako et al. (2018); SDSS DR15 Sako et al. (2018)
13291 host photoZ (KD-tree method)
0.425 ± 0.119
SN2213-1745 22:13:39.97 −17:45:24.5 SLSN-R 2.046 Cooke et al. (2012)
SN2017mf 14:16:31.00 +39:35:12.0 SN Ia 0.026 NGC 5541 14:16:31.80 +39:35:20.7 Sb 12.7 6.64 Foley et al. (2018)
SN2017yh 17:52:06.25 +21:33:58.3 SN Ia 0.020 IC 1269 17:52:05.86 +21:34:09.0 Sbc 12.0 4.86 Foley et al. (2018)
SN2013cv 16:22:43.19 +18:57:35.0 SN Ia pec 0.036 SDSS J162243.02+185733.8 16:22:43.02 +18:57:33.8 2.7 1.93 Large peak optical and UV luminosity, Cao et al. (2016b)
absence of iron absorption lines in the early
spectra
SN2006ptf 02:27:16.17 −00:23:36.5 SN Ia 0.298 SDSS J022716.08-002335.6 02:27:16.08 −00:23:35.6 1.6 7.21 According to SDSS DR15 host has BOSS Sako et al. (2018)
spectrum with z = 0.299, class = galaxy
star forming
SDSS-II SN 2661 23:32:49.80 +00:05:50.0 SN II 0.191 SDSS J233249.88+000549.2 23:32:49.88 +00:05:49.2 1.4 4.59 Sako et al. (2018)
SDSS-II SN 00:31:13.40 −00:07:08.7 ?SN II pSN II in Sako et al. (2018) Sako et al. (2018)
20266
Outliers found in three data sets with different dimensionality reduction
SDSS-II SN 21:29:40.40 −00:01:38.8 ?SN II SDSS J212940.40-000138.9 21:29:40.40 −00:01:39.0 pSN II in Sako et al. (2018); SDSS DR15 Sako et al. (2018)
12868 host photoZ (KD-tree method)
0.688 ± 0.167
SDSS-II SN 02:02:11.76 +00:13:46.3 ?SN II/?Star SDSS J020211.76+001346.2 02:02:11.76 +00:13:46.2 pSN II in Sako et al. (2018); host classified Sako et al. (2018)
19699 as star by SDSS DR15
SDSS-II SN 22:07:04.15 +00:11:00.6 ?SN Ia 22:07:04.11 +00:10:58.9 pSN II in Sako et al. (2018); host photoZ Sako et al. (2018)
16302 0.185 ± 0.015 (Smith et al. 2012)
SDSS-II SN 02:48:49.91 −00:06:27.1 ?SN Ia SDSS J024849.89-000626.6 02:48:49.89 −00:06:26.7 pSN Ia in Sako et al. (2018); SDSS DR15 Sako et al. (2018)
15745 host photoZ (KD-tree method)
0.657 ± 0.074

Downloaded from https://academic.oup.com/mnras/article/489/3/3591/5554761 by Universidad Nacional Autonoma de Mexico user on 04 June 2024
 Table A1 – continued
Name
pt α δ Typea zCMB Host name Host α Host δ Host Sep. (”)c Sep. Commentsd References
typeb (kpc)c
Gaia16ayef 19:40:01.10 +30:07:53.4 ULENS, CV MW Binary microlensing event Bakis et al. (2016)
PS1-1000007 02:23:30.71 −04:38:10.8 SN Ia 0.137 SDSS J022330.91-043810.6 02:23:30.91 −04:38:10.7 3.0 7.25 Rest et al. (2014)
SN2006ne 01:13:37.84 +00:25:26.0 SN Ia 0.046 SDSS J011337.58+002525.5 01:13:37.58 +00:25:25.5 S 3.9 3.55 According to SDSS DR15 host has BOSS Sako et al. (2018)
spectrum with z = 0.047, class = galaxy
star forming
SDSS-II SN 00:54:39.68 −00:25:23.9 SN II 0.214 SDSS J005439.88-002526.3 00:54:39.88 −00:25:26.3 3.8 13.36 Sako et al. (2018)
19504
SDSS-II SN 03:15:24.35 −00:50:54.0 ?SN II/?Star SDSS J031524.34-005053.9 03:15:24.34 −00:50:54.0 pSN II in Sako et al. (2018); host classified Sako et al. (2018)
18266 as star by SDSS DR15
PS15cfn 21:59:21.97 −21:07:10.7 91T-like 0.109 GALEXASC 21:59:22.00 −21:07:13.5 2.8 5.63 Foley et al. (2018)
J215922.04-210713.3
SDSS-II SN 02:54:57.10 −00:00:18.6 ?SN Ia pSN Ia in Sako et al. (2018) Sako et al. (2018)
15048
Outliers found in two data sets with different dimensionality reduction
SN2006ej 00:38:59.77 −09:00:56.6 SN Ia 0.020 IC 1563 00:39:00.24 −09:00:52.4 S0 8.1 3.30 In JLA (Betoule et al. 2014) and Betoule et al. (2014)
Pantheon (Scolnic et al. 2018)
cosmological samples
SDSS-II SN 02:22:42.43 +00:25:05.0 ?Unknown/?Star SDSS J022242.43+002504.8 02:22:42.43 +00:25:04.9 Unknown object in Sako et al. (2018); host Sako et al. (2018)
18391 classified as star by SDSS DR15
SN1996ai 13:10:58.13 +37:03:35.4 SN Ia 0.004 NGC 5005 13:10:56.31 +37:03:32.2 Sbc 22.0 1.82 Highly reddened SN Ia [E(B − V)host = Jha, Riess & Kirshner
1.69 ± 0.10 mag] (2007); Wang et al. (2008)
SN2016ayg 07:30:17.40 +25:01:56.0 SN Ia 0.043 SDSS J073017.25+250153.5 07:30:17.26 +25:01:53.5 3.1 2.66 Foley et al. (2018)
SDSS-II SN 4226 23:40:41.66 −00:54:21.2 ?SN II/?Star SDSS J234041.66-005421.3 23:40:41.66 −00:54:21.3 pSN II in Sako et al. (2018); host classified Sako et al. (2018)
as star by SDSS DR15
SN2005jw 20:40:19.25 −00:00:25.8 SN Ia 0.380 SDSS J204019.14-000022.8 20:40:19.14 −00:00:22.9 Sbc/Sc 3.3 17.37 In JLA (Betoule et al. 2014) and Zheng et al. (2008); Betoule
Pantheon (Scolnic et al. 2018) et al. (2014)
cosmological samples
SN2016blnf 13:34:45.49 +13:51:14.3 91T-like 0.024 NGC 5221 13:34:55.91 +13:49:57.1 Sb 170.3 82.41 Peculiar rise time, non-evolving blue Foley et al. (2018)
colour, unusual strong C II absorption
SN2016bmc 19:10:37.33 +37:39:17.0 SN Ia 0.028 UGC 11409 19:10:37.51 +37:39:18.9 S 2.9 1.61 Foley et al. (2018)
SDSS-II SN 2093 22:36:36.28 −00:12:47.0 ?SN II pSN II in Sako et al. (2018) Sako et al. (2018)
SDSS-II SN 01:29:59.18 −00:38:05.4 ?SN II SDSS J012959.31-003800.3 01:29:59.31 −00:38:00.3 Sbc zSN II in Sako et al. (2018); according to Sako et al. (2018)
17317 SDSS DR15 host has BOSS spectrum with
z = 0.118, class = star-forming galaxy
SNLS-03D1cm 02 24 55.27 −04 23 03.4 91T-like 0.870 [HSP2005] 02:24:55.28 −04:23:03.7 0.3 2.59 Peculiar Type Ia SN with the stretch-related Guy et al. (2010); Bazin
J022455.28-042303.68 parameter X1 = 4.54 et al. (2011)
SDSS-II SN 01:29:16.13 +00:42:37.9 SLSN According to table 2 of Sako et al. (2018) Sako et al. (2018)
17789 SN has four spectra
SN2015bn 11:33:41.57 +00:43:32.2 SLSN-I 0.114 SDSS J113341.53+004333.2 11:33:41.53 +00:43:33.3 1.3 2.59 Hydrogen-poor superluminous SN Le Guillou et al. (2015);
Nicholl et al. (2016)
SDSS-II SN 03:33:27.41 +00:16:10.7 ?SN II/?Star SDSS J033327.41+001610.7 03:33:27.41 +00:16:10.7 pSN II in Sako et al. (2018); host classified Sako et al. (2018)
2809† as star by SDSS DR15
Outliers found in a data set of 364 photometric characteristics (121 × 3 normalized fluxes and the LC flux maximum)
SDSS-II SN 22:45:49.70 −00:15:54.9 ?SN II pSN II in Sako et al. (2018) Sako et al. (2018)
18228
SDSS-II SN 01:22:42.61 −00:02:48.4 ?SN II pSN II in Sako et al. (2018) Sako et al. (2018)
18733
SDSS-II SN 21:43:18.71 +00:32:21.7 ?SN II SDSS J214318.74+003219.8 21:43:18.74 +00:32:19.9 zSN II in Sako et al. (2018); according to Sako et al. (2018)
19047 SDSS DR15 host has BOSS spectrum with
SNAD: SuperNova Anomaly Detection

z = 0.412, class = galaxy starburst

MNRAS 489, 3591–3608 (2019)

3607

Downloaded from https://academic.oup.com/mnras/article/489/3/3591/5554761 by Universidad Nacional Autonoma de Mexico user on 04 June 2024
 Table A1 – continued
3608

pt
Name α δ Typea zCMB Host name Host α Host δ Host Sep. (”)c Sep. Commentsd References
typeb (kpc)c
SDSS-II SN 02:44:37.90 +00:46:32.1 ?SN II/QSO SDSS J024437.89+004631.9 02:44:37.89 +00:46:32.0 pSN II in Sako et al. (2018); according to Sako et al. (2018)
19395 SDSS DR15 host has BOSS spectrum with
z = 1.318, class = QSO
SDSS-II SN 6992 22:10:25.09 +00:00:02.7 ?SN II/?Star SDSS J221025.08+000002.5 22:10:25.08 +00:00:02.5 pSN II in Sako et al. (2018); host classified Sako et al. (2018)
as star by SDSS DR15
SN2005mp 01:04:45.68 +00:03:20.2 SN Ia 0.272 Host identified by Sako et al. 2018 (SDSS Ganeshalingam, Li &
J010445.51+000320.8) has BOSS Filippenko (2013); Sako
spectrum with z = 0.952 that is different et al. (2018)

MNRAS 489, 3591–3608 (2019)

from SN redshift
SN2013abe 14:32:44.49 +09:53:12.3 SN IIP 0.006 NGC 5669 14:32:43.80 +09:53:28.8 Scd 19.4 2.40 The LC and spectra suggest that the SN is a Bose et al. (2015a)
normal Type IIP event, although with a
steeper decline during the plateau relative to
other archetypal SNe of similar brightness
SN1999gi 10:18:16.66 SN IIP 0.002 NGC 3184 10:18:16.99 Sc 60.5 2.51 Leonard et al. (2002)
M. V. Pruzhinskaya et al.

+41:26:28.2 +41:25:27.8
Outliers found in a data set of 10 Gaussian process parameters (9 fitted parameters of the kernel and the log-likelihood of the fit)
PTF10aagc 09:39:56.93 +21:43:16.9 SLSN-I 0.207 SDSS J093956.91+214317.1 09:39:56.91 +21:43:17.1 0.3 1.16 SLSN-I with hydrogen in late spectra; host Perley et al. (2016); De Cia
morphological structure suggests a possible et al. (2018)
ongoing merger
SN2006T 09:54:30.21 −25:42:29.3 SN IIb 0.008 NGC 3054 09:54:28.61 −25:42:12.4 Sbc 27.4 4.51 Monard (2006); Stritzinger
et al. (2018)
SN2010bb 10:44:38.23 +57:48:40.0 SN Ia 0.118 SDSS J104438.19+574839.8 10:44:38.19 +57:48:39.8 0.4 0.80 Rest et al. (2014)
SN2013ej 01:36:48.16 +15:45:31.0 SN IIP/SN IIL 0.002 NGC 628 01:36:41.77 +15:47:00.5 Sc 128.5 5.32 LC is intermediate between those of Type Bose et al. (2015b); Huang
IIP and IIL SNe et al. (2015); Mauerhan et al.
(2017)
SNLS-04D2gb 10:02:22.67 +01:53:39.0 SN Ia 0.452 SDSS J100222.66+015339.2 10:02:22.66 +01:53:39.2 0.2 1.44 In JLA (Betoule et al. 2014) and Betoule et al. (2014)
Pantheon (Scolnic et al. 2018)
cosmological samples; host classified as
star by SDSS DR15
SNLS-05D2mp 09:59:08.63 +02:12:14.4 SN Ia 0.355 In JLA (Betoule et al. 2014) and Betoule et al. (2014)
Pantheon (Scolnic et al. 2018)
cosmological samples
SNLS-06D2ag 10:01:43.36 +01:51:37.1 SN Ia 0.310 SDSS J100143.26+015135.4 10:01:43.26 +01:51:35.4 2.3 10.33 Guy et al. (2010);
Ganeshalingam et al. (2013)
SNLS-06D3cn 14:19:25.85 +52:38:27.5 SN Ia 0.232 SDSS J141925.79+523825.9 14:19:25.79 +52:38:25.9 1.7 6.25 Guy et al. (2010)
SN1999cc 16:02:42.03 +37:21:34.4 SN Ia 0.032 NGC 6038 16:02:40.55 +37:21:34.2 Sbc 17.6 11.28 In JLA (Betoule et al. 2014) and Betoule et al. (2014)
Pantheon (Scolnic et al. 2018)
cosmological samples
SN2002aw 16:37:29.06 +40:52:50.3 SN Ia 0.026 2MFGC 13321 16:37:29.22 +40:52:48.2 Sb 2.8 1.45 Ganeshalingam et al. (2013)
SN2002eb 22:19:05.24 +24:35:39.8 SN Ia 0.026 PGC 68560 22:19:06.29 +24:35:53.4 Sb 19.7 10.33 Ganeshalingam et al. (2013)
SN2004dt 02:02:12.77 −00:05:51.5 SN Ia 0.019 NGC 799 02:02:12.30 −00:06:02.6 Sa 13.1 5.07 Spectral subtype: high velocity (HV; Wang Folatelli et al. (2013)
et al. 2009), broad line (BL, Branch et al.
2006)
SN2009bw 03:56:06.92 +72:55:40.9 SN IIP 0.004 UGC 2890 03:56:04.44 +72:55:18.5 Sdm 24.9 2.06 Luminosity drop from the photospheric to Inserra et al. (2012);
the nebular phase is one of the fastest ever Rodrı́guez, Clocchiatti &
observed, 2.2 mag in 13 d Hamuy (2014)
Notes: a Type of the source. A prefix “?” means that the source is not confirmed spectroscopically.
b Simbad host galaxy morphological type.
c Separation of the source from the center of its host galaxy.
d If classification is made by Sako et al. 2018: a prefix “p” (pSN) indicates a purely photometric type, a prefix “z” (zSN) indicates that a redshift is measured from its candidate host galaxy and the classification uses that redshift as a prior.
e The object is also found in a data set of 364 photometric characteristics (121 × 3 normalized fluxes and the LC flux maximum).
f The object is also found in a data set of 10 Gaussian process parameters (9 fitted parameters of the kernel and the log-likelihood of the fit).

This paper has been typeset from a TEX/LATEX file prepared by the author.

Downloaded from https://academic.oup.com/mnras/article/489/3/3591/5554761 by Universidad Nacional Autonoma de Mexico user on 04 June 2024