Classifying Broad Absorption Line Quasars: Metrics, Issues and a New Catalogue Constructed from SDSS DR5
aa r X i v : . [ a s t r o - ph . C O ] J u l Mon. Not. R. Astron. Soc. , 1–10 (2009) Printed 1 November 2018 (MN L A TEX style file v2.2)
Classifying Broad Absorption Line Quasars: Metrics,Issues and a New Catalogue Constructed from SDSS DR5
S. Scaringi ⋆ , C.E. Cottis , C.Knigge , M.R. Goad Department of Physics and Astronomy, University of Southampton, Highfield, SO17 1BJ, UK Department of Physics and Astronomy, University of Leicester, University Road, LE1 7RH, UK
ABSTRACT
We apply a recently developed method for classifying broad absorption line quasars(BALQSOs) to the latest QSO catalogue constructed from Data Release 5 of theSloan Digital Sky Survey. Our new hybrid classification scheme combines the powerof simple metrics, supervised neural networks and visual inspection. In our view theresulting BALQSO catalogue is both more complete and more robust than all previousBALQSO catalogues, containing 3552 sources selected from a parent sample of 28,421QSOs in the redshift range 1 . < z < .
2. This equates to a raw BALQSO fraction of12 . Key words: quasars: absorption lines, methods, catalogues, surveys
Broad absorption line quasars (BALQSOs) are a sub-class of active galactic nuclei (AGN) exhibiting strong,broad and blue-shifted spectroscopic absorption features(Foltz et al. 1990; Weymann et al. 1991; Reichard 2003b;Hewett & Foltz 2003). These features are thought to beformed in fast (0 . c -0 . c ) and powerful outflows from the ac-cretion disk around the supermassive black hole at the heartof the AGN (Korista 1992). The vast majority of BALQSOsare radio-quiet (Stocke et al. 1992, but see Brotherton et al.2006 for some counter examples), and there are subtle dif-ferences between their continuum and emission line prop-erties and those of normal (non-BAL) QSOs (Reichard2003b). However, despite these differences, BALQSOs andnon-BALQSOs appear to be drawn from the same parentpopulation (Reichard 2003b). Most BALQSOs belong to ⋆ E-mail: [email protected] the subclass of the so-called HiBALs, only displaying ab-sorption in certain high-ionisation lines (e.g. N v iv iv ii all QSOs may undergo significant mass lossthrough winds (Ganguly & Brotherton 2007), but BALs areonly observed if the central continuum and/or emission linesource is viewed directly through the outflowing material.Viewed in this context, BALQSOs may be the only availabletracers of a key physical process common to all AGN. Also,the powerful outflows we observe in BALQSOs are an im-portant example of AGN feedback in action (Tremonti et al.2007). Such feedback is a key ingredient required in theoret-ical attempts to understand galaxy “downsizing” and mayalso be responsible for regulating the growth of supermas-sive black holes. Moreover, the fraction of QSOs displaying c (cid:13) S. Scaringi et al.
BAL features ( f BALQSO ) may provide a direct estimate ofthe opening angle of these outflows.Historically, BALQSO samples have been selectedon the basis of the so-called balnicity index (BI;Weymann et al. 1991) or similar metrics. These samplesconsistently yielded BALQSO fraction estimates in therange f BALQSO ≈ .
10 - 0 .
15 (Weymann et al. 1991;Tolea et al. 2002; Hewett & Foltz 2003; Reichard 2003a). Ina previous paper (Knigge et al. 2008; Paper I), we showedthat both the BI and a more recently defined metric, theabsorption index (AI; Trump et al. 2006), are biased whenselecting BALQSOs, the former being incomplete at the low-velocity end of the BALQSO distribution, and the latter suf-fering from significant contamination by objects with low-velocity absorption systems which may be unrelated to thehigher velocity outflows.Here, we use a combination of the classic BI met-ric, a simple neural network and visual inspection (thehybrid-LVQ approach we developed in Paper I) to pro-duce a BALQSO sample that is both more completethan purely BI-based ones and, importantly, signifi-cantly more robust than AI-based ones. We have ap-plied our hybrid-LVQ algorithm to the QSO sample as-sociated with Data Release 5 (DR5) of the Sloan Dig-ital Sky Survey (SDSS;Adelman-McCarthy et al. 2007;Schneider et al. 2007) using the BIs calculated fromGibson et al. (2009). The resulting catalogue contains 3552BALQSOs selected from a parent sample of 28,421 QSOson the basis of absorption close to the C iv high-ionisation emission line. This catalogue may be obtainedfrom ∼ simo . A prelimi-nary version of the catalogue has already been presented inScaringi et al. (2008). In addition, we also provide (at thesame address) a catalogue of the meta-data, i.e. the datapertaining to the parent QSO sample and subsequently usedin the compilation of our BALQSO catalogue, so that mem-bers of the scientific community wishing to compile theirown BAL/non-BAL subsamples may readily do so. The SDSS DR5 QSO catalogue contains over 77,000 objectsin total (Schneider et al. 2007). However, for the purposeof constructing a uniform BALQSO catalogue, we only con-sider objects whose spectra fully cover the C iv − blueward of the C iv line centre), which displays aparticularly deep and well-defined absorption trough in thespectra of most BALQSOs. Given the wavelength range cov-ered by the SDSS spectra, this implies an effective redshiftwindow of 1 . < z < . Our BALQSO classification method works on continuumnormalised spectra covering the wavelength range 1401 ˚A - 1700 ˚A with 1 ˚A dispersion. It also uses the associated BIsfor training the neural network and to flag borderline casesrequiring visual inspection. The BI metric is defined as BI = − Z (cid:20) − f ( v )0 . (cid:21) Cdv. (1)Here, the limits of the integral are in units of km s − , f ( v )is the normalised flux as a function of velocity displacementfrom line centre. The constant C = 0 everywhere, unlessthe normalised flux has satisfied f ( v ) < . − , at which point it is switched to C = 1 until f ( v ) > . BI > − .The BI by definition excludes strong, low-velocity absorp-tion systems; for example, any deep absorption of width3000 km s − which starts less than 2000 km s − blueward ofthe rest wavelength of the C iv emission line will be assigned BI = 0 km s − . Thus BALQSO catalogues constructed us-ing the BI metric are likely to be significantly incomplete atthe low velocity end of the distribution.For this reason Hall et al. (2002) introduced the so-called AI (Absorption Index), in an attempt to recover thoselow-velocity absorption systems objects that were missed bythe BI. The AI is defined as AI = Z [1 − f ( v )] Cdv, (2)here now C = 1 in all regions where f ( v ) > . − and C = 0 otherwise. The twokey differences that allow objects with BI = 0 km s − toachieve AI > − are (i) that the AI includes regionswithin 3000 km s − of line centre (and also regions beyond25 ,
000 km s − ) and (ii) that the AI includes objects withmuch narrower absorption troughs than the BI. The remain-ing differences are associated with the presence [absence] ofthe factor 0 . . More specifically, itshows the average properties of QSO spectra on a grid in We have used the DR3 subset so that we can use the AIs pro-vided by Trump et al. (2006) c (cid:13) , 1–10 lassifying BALQSOs: Metrics, Issues and a New catalogue AI/BI space, allowing a close examination of the absorptionthrough dependence on the AI and the BI. For referencewe have also included in each panel the same non-BALQSOcomposite created from QSOs with AI = 0 km s − (dashedgreen curve).It is generally clear from Fig. 1 that both the AI andthe BI tend to select redder QSOs, and that not only do thetroughs get wider with increasing AI/BI, but also deeper.Moreover, Fig. 1 shows how QSO samples selected fromthe low-velocity region of the AI do not display the “tra-ditional” BALQSO properties. This is best shown in thesecond composite from the left panel (top) with BI = 0km s − and 1 km s − AI <
500 km s − , which shows little,if any, sign of absorption when compared to the non-BALcomposite. The next panel down displays a composite cre-ated from 1561 QSOs (the largest sub-sample) which have500 km s − < AI < − and BI = 0 km s − . Rel-ative to the non-BAL composite, there is some evidence ofabsorption close to the C iv emission line. However, we cau-tion that, since, the BAL composite is significantly redderthan the non-BAL composite, identifying broad absorptionlines in this spectrum is difficult, without first dereddeningthe spectrum. The remaining panels show spectra with in-creasingly prominent absorption in the vicinity of C iv , withthe absorption strength (depth) and reddening increasingboth with increasing AI (moving from top to bottom), andwith increasing BI (left to right).We conclude from examining Fig. 1 that using the AI > − to select BALQSOs is unreliable, sinceQSOs with 0 km s − < AI < − and BI = 0km s − have spectra that are not different from AI = 0km s − non-BALQSOs. Moreover, most BALQSOs fall inthe low-velocity region of both the AI and the BI contin-uum, which is also the region which turns out to be thehardest to classify. However, it is interesting to note thatQSOs with AI > − and BI <
500 km s − dolook like BALQSOs. We have decided to omit the AI metricfrom our hybrid classification method since about ≈
50% ofthe objects selected using this metric may not be genuineBALQSOs (see Paper I). Instead, our hybrid-LVQ methoduses the BI, a simple neural network and visual inspectionto select BALQSOs.
The method we use to classify BALQSOs has already beendescribed in detail in Paper I, so we only provide an overviewof the key points here. Briefly, our method is a hybridof BI-based, neural network and visual classifications, andis designed to produce a more complete BALQSO samplethan a pure BI selection, but without significantly increas-ing the number of false positives. Starting with a BI-basedclassification (as calculated by Gibson et al. 2009), we usea simple neural network-based machine learning algorithmcalled “learning vector quantization” (LVQ, Kohonen 2001)to identify objects that might have been miss-classified bythe BI. All such objects are then inspected and classifiedvisually.We caution that both the measured BI (and the AI forthe reference) are very sensitive to our ability to perform anaccurate fit to the underlying continuum. Overestimatingthe underlying continuum strength can yield a large posi- tive AI and BI in the absence of any absorption. Conversely,if the continuum is underestimated, weak broad absorptionfeatures may go unrecognised. This is an issue which can alsoaffect our hybrid-LVQ classification method. For this reasonwe have decided to use the BIs calculated from Gibson et al.(2009), since their continuum fitting algorithm is likely tobe superior to the one we use for normalising spectra forinput into LVQ. This is mainly because they employ mul-tiple composites in order to fit the underlying continuum(Trump et al. 2006).For input into LVQ, we normalise all QSO spectrausing the method described in Knigge et al. (2008) andNorth et al. (2006), in which each spectrum is fitted witha modified DR3 QSO composite (constructed from objectswith AI = 0 km s − as calculated from Trump et al. 2006)allowing for object-to-object differences in reddening andspectral index. We then bin each spectrum onto a uniformgrid in wavelength, and use the binned spectrum between1401˚A - 1700˚A for our classification purposes.The way we train our LVQ network to recognize BALQ-SOs has been described in detail in Paper I. In brief, we em-ploy a training set composed of 400 BI > − and 400 BI = 0 km s − QSOs and train our LVQ-network to recog-nise
BI > − objects at first. We then visually inspectour neuron map for BALQSO mis-classifications (locating BI > − QSOs in BI = 0 km s − nodes and viceversa) and re-tag those objects. We then retrain our LVQ-network using the new BALQSO vs. non-BALQSO tags tocreate a final neuron map. Note that redshift uncertaintiesare explicitly taken into account by our network and all thespectra have been de-reddened to match the non-BALQSOcomposite. Below, we will sometimes refer to the full hybridmethod as LVQ-based, but it is always worth keeping inmind that LVQ is only one part of a process also involvingthe BI and visual inspection. Our LVQ based DR5 BALQSO catalogue contains 3552 ob-jects ( ≈ .
5% of the parent QSO sample). Fig. 2 shows aflow diagram detailing the individual steps involved in creat-ing this catalogue, along with the numbers of QSOs associ-ated with each step. Overall, we find that 3205 QSOs (11 . BI > − ), and 3282 QSOs (11 . . . .
0% (712/3552) of the objects in our final catalogue andproduced false positives at a rate of 11 .
4% (365/3205). Simi- c (cid:13) , 1–10 S. Scaringi et al.
Figure 1.
Composites in various AI-BI ranges (blue line), and composites created from AI = 0 km s − and BI = 0 km s − objects(dashed green line). The AI and BI bins on the side panels are in km s − . Reddening has not been taken into account.c (cid:13) , 1–10 lassifying BALQSOs: Metrics, Issues and a New catalogue Figure 2.
Flow diagram illustrating the steps involved in ourhybrid-LVQ classication method. larly, LVQ alone would have missed 20 .
0% (710/3552) of ourBALQSOs and produced false positives at a rate of 13 . iv line shows no sign of absorption, andthus this object was not classified as a BALQSO by us. How-ever, there is some evidence that the Lyman alpha line does show reasonable broad, blue-shifted absorption. By contrast,the spectra on the right have BI = 0, despite the fact thatthey show signs of absorption (and are therefore included inour catalogue). These objects were recognized by the LVQnetwork and, due to the disagreement between the BI andLVQ verdicts, visually inspected for final classification. One last note of caution concerns the rates of false pos-itives and negatives among objects that were not visuallyinspected. While the sample of 2227 BALQSOs that were inspected visually may be considered to be fairly reliable,the samples of non-inspected objects are not as clean. Inparticular, since LVQ and BI alone produce false positivesat rates of 11 .
4% and 13 . .
5% (0 . × . . × .
2) of true BALQSOs. This amounts toroughly 85 false negatives, i.e. 85 BALQSOs that are missingfrom our catalogue.Because of the many problems encountered when tryingto compile BALQSO catalogues, we have decided to pro-duce for the scientific community a meta-catalogue whichincludes our whole DR5 parent sample used in this work in-stead of just a BALQSO catalogue . The first ten entries ofthis meta-catalogue are presented in Table 1. For each QSOin our parent sample we provide all tags that we have foundto be useful in creating our own hybrid-LVQ BALQSO cat-alogue. In this section we highlight the difficulties in compilingBALQSO samples using composites derived from our QSOmeta-catalogue as shown in Fig. 4. Each panel displaysthe non-BALQSO composite (shown in dashed green) nor-malised to 1750 ˚A and reddened to match the other com-posites shown in each panel (solid blue lines), which werecreated by selecting relevant QSO sub-sets culled from themeta-catalogue. In the top, row we show composites fromQSOs in our parent sample which were finally classified asBALQSOs by our hybrid-LVQ method, subdivided into ob-jects with AI = 0 km s − (top-left), BI = 0 km s − (topmiddle) and LVQ non-BAL. All of these composites showclear signatures of absorption blueward of C iv . We notethat, for consistency, we have only used objects already in-cluded in SDSS DR3 in constructing the composites shownin the left panels, since only these have AI values calculatedby Trump et al. (2006).The composite in the upper left panel comprises ob-jects with AI = 0 km s − (and therefore also BI = 0km s − ) that were classified as BALQSO by the LVQ net-work and subsequently confirmed as BALQSOs visually. Al-though only 126 objects were used for the creation of thiscomposite, the absorption near C iv and the slightly trun-cated emission line are clear BALQSO signatures. Theseare mostly BALQSOs whose troughs are smaller that 1000km s − (and hence with AI = 0 km s − ).The BALQSO composite in the top middle panel wascreated from objects with BI = 0 km s − , but subsequently An electronic version can be found at and on the VizieRserverc (cid:13) , 1–10
S. Scaringi et al.
Figure 3.
Four QSO spectra with different classification tags from the Gibson et al. (2009) catalogue and ours. The two spectra on theleft have positive BI’s but are not included in our BALQSO catalogue, whilst the two spectra on the right have BI = 0 km s − and areincluded in our catalogue. Table 1.
First 10 objects from our DR5 meta-catalogue. The column names are the same as those used by the SDSS team with theexception of the last 2.
LVQ tag is set to 1 if the neural network regarded the QSO as a BALQSO, 0 of not.
Final tag is set to 1 ifthe QSO is considered as a BALQSO by our hybrid-LVQ method. The BIs have been taken from Gibson et al. (2009). The ts t qso and ts t hiz columns represent Low-z Quasar selection flag and High-z Quasar selection flag respectively as defined by the SDSS team(Schneider et al. 2007). The catalogue can be found in electronic format from and the VizieRserver. SDSS Name RA DEC z
M i ts t qso ts t hiz BI LVQ tag Final tag deg. deg. M i km s − − − − − − (cid:13) , 1–10 lassifying BALQSOs: Metrics, Issues and a New catalogue Figure 4.
Various composites created in order to examine our BALQSO classification parameter space. The solid blue line displayscomposites created by selecting different QSOs in our classification parameter space. The dashed green line is a composite created with AI = 0 km s − QSOs after being normalised at 1750˚A and de-reddened to match the composites in each panel.c (cid:13) , 1–10
S. Scaringi et al. identified as BALQSOs by the LVQ neural network. Thiscomposite shows the largest amount of reddening and afairly narrow and deep absorption blueward of C iv . Suchnarrow features will, by definition, be classified as non-BALsusing the BI metric since they lie within 3000 km s − of theline centre.. In the upper right panel we show the compos-ite formed from QSOs with BI > − , identified asnon-BALs by the LVQ network, but finally included in thecatalogue on the basis of visual inspection. This compositeshows a strong, broad, absorption line blueward of C iv , butvery little reddening. Furthermore, since the composites inboth the top right and top middle panels contain similarnumbers of QSOs, it is evident that neither method alone isas reliable in identifying BALQSOs as one might hope. Allof the information needed to recreate these composites canbe found in our meta-catalogue by querying on various tags.The bottom row of Fig. 4 compares those compositescomprising QSOs classified as BALQSO candidates usingonly a single metric that were then re-classified as non-BALsafter visual inspection. In detail, the composite at bottomleft is created from objects with AI > − , but finallyclassified as non-BALs by our hybrid-LVQ method. There isvirtually no evidence for absorption in this composite spec-trum, despite all of the 2362 comprising this composite hav-ing AI > − . This again points out the problematicnature of the AI metric for BALQSO selection purposes.The middle bottom panel displays a composite createdfrom objects having BI > − (and therefore also AI > − ) and a non-BALQSO LVQ tag, finally clas-sified by us as a non-BALQSO. This composite looks verysimilar to the one on the top right, which was already dis-cussed above. Here again, we see a fairly strong, smooth andbroad ( > − ) absorption feature associated withC iv . The similarity between these two composites is not en-tirely unexpected, since all of the objects forming them hadthe same automated classifications (positive BI and a non-LVQ tag), and thus differed only in the outcome of the visualinspection step. Since disagreement between BI and LVQ ismost likely to happen for difficult borderline cases, we shouldcertainly expect some mis-classifications and thus overlapbetween the two sub-sets of QSOs represented by these com-posites. However, closer inspection does reveal some signif-icant differences between the composites that point to thesubtle, but consistent absorption line properties that wereobviously picked by the visual classification step. For exam-ple, the peaks of the C iv and Lyman- α lines are lower in thetop right composite than in the bottom middle one, and onlythe top right one shows clear evidence of absorption eatinginto the blue wing of the C iv line (compared to the non-BAL composite). Moreover, even though both compositesshow some evidence for absorption affecting the bluest partof the spectrum – shortward of Lyman alpha, and particu-larly around the Lyman beta and O vi blend near 1030 ˚A –this absorption is stronger in the top right panel. Finally, thebroad absorption trough associated with C iv in the bottommiddle panel is suspiciously symmetric between 2000 km s − and 20 ,
000 km s − , the limits within which the BI is calcu-lated. This may indicate that this trough is formed from thesuperposition of many narrow lines that may or may not beassociated with the traditional BAL-flow.All of these differences are consistent with the idea thatthe objects represented in the top right panel (which is in- cluded in our final BALQSO catalogue) are more likely tobe genuine BALQSOs than those represented in the bot-tom middle panel (which are not included in our final cata-logue). However, there is no escaping the fact that the dif-ferences are extremely subtle and that a definitive classifica-tion scheme for such borderline cases remains elusive. Thisconclusion is supported by a visual re-inspection of all ofthe objects contained in these two sub-samples: while wegenerally remain happy with our classifications as ”best-betestimates”, it is clear that in many cases a definitive classi-fication is impossible. Since the 365 borderline cases repre-sent 11% of the Gibson et al. (2009) sample, we caution thatthere is a systematic uncertainty of ∼
11% on the BALQSOfraction suggested by even the best presently available clas-sification schemes. This is one of the key reasons we havedecided to provide the community with all of the meta-datawe have used in constructing our own catalogue.The last figure on the bottom right displays a compositecreated by selecting QSOs originally classified as BALQSOsby our neural network but re-classified as non-BALQSOsduring the visual inspection phase (these objects all had BI = 0 km s − by definition or they would have not beeninspected visually). The composite here is somewhat red-der than the non-BAL (E(B - V ) = 0.04), but no clearsignatures of absorption are present. This highlights theimportance of a visual inspection phase when constructingBALQSO catalogues.To summarise, it is clear that no single metric (or vi-sual intervention) is adequate in deriving both complete andclean samples of BALQSOs at the moment, so a varietyof complementary metrics should instead be employed. Ourown experience with unsupervised and supervised learningnetworks shows that, even though much of the classificationwork may indeed be automated, human intervention is notonly useful, it is often a necessity when dealing with classi-fication involving not so clearly defined training samples. Fig. 5 shows f BALQSO as a function of signal-to-noise for BI-selected QSOs, LVQ selected QSOs and our final BALQSOfraction using our hybrid-LVQ method. The same trend asthat found by Gibson et al. (2009) is evident for BI selectedBALQSOs: f BALQSO steadily increases from ≈
9% in lowsignal-to-noise data up to 15% in high signal-to-noise data.We suspect that this is because in low signal-to-noise dataeven relatively small random fluctuations in a shallow BALtrough can trigger the zero reset in the BI calculation andcan thus result in BI = 0 km s − . We note that this wouldnot necessarily be the case if BALs were identified using amore sophisticated metric than the BI to isolate the BAL.We cannot rule out, however, that the apparent trend inthe BAL fraction with S/N has a more interesting cause,such as an underlying trend with redshift or luminosity (i.e.the BAL fraction may be higher among low-redshift and/orhigh-luminosity QSOs, which would also have higher S/Nspectra, on average). However, the simpler and more mun-dane explanation – that the trend is primarily due to thedifficulty in identifying BAL features in low-S/N spectra –seems far more likely. We have also visually inspected someof the objects with high BI that are not included in our fi-nal catalogue and conclude that these are cases where the c (cid:13) , 1–10 lassifying BALQSOs: Metrics, Issues and a New catalogue
10 202 3 4 5 6 7 8 9 30 400.080.10.120.140.16 S/N f BA L Q S O BALsBIsLVQ
Figure 5. f BALQSO as a function of signal-to-noise calculatedin the same way as Gibson et al. (2009) for BI-selected QSOs,hybrid-LVQ selected QSOs and the final BALQSOs included inour final hybrid-LVQ catalogue.
BI must have been calculated incorrectly and should havebeen set to BI = 0 km s − .By contrast, the BALQSO fraction produced by LVQalone at high S/N levels is roughly constant, and slightlylower than the fraction suggested by the BI or indicated byour final catalogue. Thus the maximum efficiency of LVQ(when working on high-quality spectra) is comparable to,but slightly less, than that of the BI. Fig. 5 also shows thatthe LVQ-suggested BALQSO fraction actually increases to-wards the lowest S/N levels. Given that the number of low-S/N BALQSOs suggested by LVQ alone is actually higherthan that in our final catalogue, and that every LVQ-selectedBALQSO candidate was either included in the catalogue orrejected as a false positive via visual inspection, this im-plies that LVQ has a tendency to classify low-S/N spectraas BALQSOs, leading to a higher false positive rate in thislimit. This is not entirely unexpected and actually meansthat LVQ and BI selections are highly complementary meth-ods when applied across the full range of S/N levels. We have used a recently developed technique for identifyingbroad absorption lines in quasar spectra to compile a morerobust and complete BALQSOs catalogue. Our technique isbased on a combination of the traditional “balnicity index”,a simple neural network and visual inspection of borderlinecases and is designed to produce BALQSO samples that aremore complete than purely BI-based ones, while still avoid-ing a high incidence of false positives. Our final cataloguecovers the redshift range 1 . < z < . ≈ .
5% ofthe SDSS DR5 QSOs parent sample with a false positive rateof ∼ ACKNOWLEDGEMENTS
REFERENCES
Adelman-McCarthy J. K., Ag¨ueros M. A., Allam S. S., An-derson K. S. J. e. a., 2007, ApJS, 172, 634Brotherton M., de Breuck C., Schaefer J., 2006, MNRAS,pp L90+Foltz C. B., Chaffee F. H., Hewett P. C., Weymann R. J.,Morris S. L., 1990, in Bulletin of the American Astronom-ical Society Vol. 22 of Bulletin of the American Astronom-ical Society, On the Fraction of Optically-Selected QSOswith Broad Absorption Lines in Their Spectra. pp 806–+Ganguly R., Brotherton e. a., 2007, ApJ, 665, 990Gibson R. R., Jiang L., Brandt W. N., et al. H., 2009, ApJ,692, 758 c (cid:13) , 1–10 S. Scaringi et al.
Hall P. B., Anderson S. F., Strauss M. A., York D. G. a. a.,2002, ApJS, 141, 267Hewett P., Foltz C., 2003, AJ, 125, 1784Hogg D. W., Lang D., 2008, in American Institute ofPhysics Conference Series Vol. 1082 of American Insti-tute of Physics Conference Series, Astronomical imaging:The theory of everything. pp 331–338Knigge C., Scaringi S., Goad M. R., Cottis C. E., 2008,ArXiv e-prints, 802Kohonen T., 2001, Self-organizing maps. Self-organizingmaps.3rd ed.Berlin: Springer, 2001, xx, 501 p.Springer se-ries in information sciences, ISBN 3540679219Korista K. e. a., 1992, ApJ, 401, 529North M., Knigge C., Goad M., 2006, MNRAS, 365, 1057Reichard T. e. a., 2003a, AJ, 125, 1711Reichard T. e. a., 2003b, AJ, 126, 2594Scaringi S., Cottis C. E., Knigge C., Goad M. R., 2008, inAmerican Institute of Physics Conference Series Vol. 1082of American Institute of Physics Conference Series, BroadAbsorption Line Quasar catalogues with Supervised Neu-ral Networks. pp 191–195Schneider D. P., Hall P. B., Richards G. T., et al. S., 2007,AJ, 134, 102Stocke J., Morris S., Weymann R., Foltz C., 1992, ApJ,396, 487Tolea A., Krolik J. H., Tsvetanov Z., 2002, ApJ, 578, 31Tremonti C. A., Moustakas J., Diamond-Stanic A. M.,2007, ApJL, 663, L77Trump J. R., Hall P. B., Reichard T. A., Richards G. T.e. a., 2006, ApJS, 165, 1Weymann R., Morris S., Foltz C., Hewett P., 1991, ApJ,373, 23 c (cid:13)000