Well most of those would require grid search tuning of at least one or
two hyperparameters to be able to compare. For instance (kernel SVR
needs at least C and gamma when the RBF kernel is used). Some would
also require scaling the input data while others do not. That would
probably be too costly to run the full grid search as an example.

I would be possible to forge a list of toy 2D yet informative datasets
(linearly separable / regular vs. highly hilly / irregular) , run the
grid search to find the optimal hyperparams for each model and then
build an example with plots such along with the optimal hyperparams
hardcoded:

http://scikit-learn.org/dev/auto_examples/plot_classifier_comparison.html
http://scikit-learn.org/dev/auto_examples/cluster/plot_cluster_comparison.html#example-cluster-plot-cluster-comparison-py

As a rule of thumbs, I would recommend that if you don't know which
one to use, just focus on RidgeCV for small to medium datasets,
SGDRegressor for very large linear datasets (e.g. more than 100k
samples) and also GBRT and take the one that works the best on your
data after some grid search for their main hyperparameters (read the
doc). Better than taking the best, averaging the predictions will
probably yield even better scores.

If you have special requirements such as sparse solution you probably
already know that you need some sort of L1 penalized model.

--
Olivier

------------------------------------------------------------------------------
Monitor your physical, virtual and cloud infrastructure from a single
web console. Get in-depth insight into apps, servers, databases, vmware,
SAP, cloud infrastructure, etc. Download 30-day Free Trial.
Pricing starts from $795 for 25 servers or applications!
http://p.sf.net/sfu/zoho_dev2dev_nov
_______________________________________________
Scikit-learn-general mailing list
[email protected]
https://lists.sourceforge.net/lists/listinfo/scikit-learn-general

Reply via email to