Repository for exploring machine learning algorithms for quasar selection/classification. These are intended for analysis of SDSS+SpIES data, but can serve as an introduction for analysis of other data sets (e.g., DES, LSST, etc.) for those looking for a template to follow. These are largely "sand boxes", so apologies if they aren't fully coherent. Though I have gone through them to try to make sure that they have some longer-term value.
SpIESHighzCandidateSelection.ipynb
is a notebook that takes the training sets that we used for KDE
selection of 3.5<z<5 quasars from SDSS+SpIES data and does color-based
self classifation using other machine learning algorithsm from
Scikit-Learn. KDE was about 78% complete and 97% efficient. The
sklearn algorithms were about the same, which gives confidence that
the sample is not highly contaminated (which isn't to so that it
isn't contaminated at all) and that we could replace our usage of
the proprietary KDE algorithm with public algorithms from sklearn. I
looked at neural networks, SVM, random forests, bagging, and boosting.
While bagging has the highest efficiency, it also takes considerably
longer than random forests, which provided the best overall solution
(in terms of completeness, efficiency and computing time). The
results of this get applied to the test set in
SpIESHighzQuasars.ipynb.
SpIESHighzCandidateSelection2.ipynb
is similar to the notebook above, but here I was exploring what would
happen if we added i-band magnitude and u-band extinction as
classification attributes. I also explored some different
combinations of selection algorithms. I found that bagging including
i and extinctu has less contamin than anything else.
SpIESHighzQuasars2.ipynb is the notebook where I explore these
changes in selection.
SpIESHighzQuasars.ipynb
is a notebook that applies the RF classifier from
SpIESHighzCandidateSelection.ipynb to the test set. This uses
colors only. I also explored with bagging, but that crashes and later
led to trying just to classify Stripe 82 objects with stand-alone
python code, running on the server. See SpIESHighzQuasarsS82all.py.
But this notebook does successfully produce output classifications
for the test set using the RF classifier. The output is in
GTR-ADM-QSO-ir_good_test_2016_out.fits.
SpIESHighzQuasars.ipynb
is a notebook that applies the SVM and bagging classifiers from
SpIESHighzCandidateSelection2.ipynb to the test set. I also tried
to create a new version of the test set that included i-band magnitude
and u-band extinction, but I ran into all sorts of problems. (Though
it may be worth trying again). In the end, I ended up writing a
stand-alone script SpIESHighzQuasarsS82all.py that just classifies
the objects in Stripe 82. However, it does so using RF, SVM, and
bagging. That way we can compare the results and find out which
produces the least contamination as that is important for our
clustering analysis.
SpIESHighzQuasarsS82all.py
See note above. Outputs GTR-ADM-QSO-ir_good_test_2016_out_Stripe82all.fits.
SpIESHighzQuasarPhotoz.ipynb
The above notebooks just do selection, now we need to do photo-$z$ on
the output files with the quasar canddiates. Here I play with
Nadaraya-Watson, Random Forest, Gaussian Process Regression, SVM, and
Stochastic Gradient Descent. The first two are good, the next two
crash, and the last one is terrible. Just doing self classification
of the training set here.
SpIESHighzQuasarPhotoz2.ipynb
Now we apply the best photo-z algorithms to the test data,
specifically quasar candidates. Right now just sticking to the Stripe
82 quasar candidates from
GTR-ADM-QSO-ir_good_test_2016_out_Stripe82all.fits as the
Nadaraya-Watson algorithm isn't particularly speedy. Output is
GTR-ADM-QSO-ir_good_test_2016_out_Stripe82all_zphot.fits.
Included quasar candidates file, which includes all of the test
object, not just the quasar candidates:
GTR-ADM-QSO-ir_good_test_2016_out_Stripe82all.fits.bz2
Actually, I couldn't because it exceeded the file size limit. So ask me if you want it!
Included the quasar candidate file limited to quasar candidates only
with two photo-z outputs:
GTR-ADM-QSO-ir_good_test_2016_out_Stripe82all_zphot.fits
This work was supported in part by NSF CDS&E grant 1411773.