Disable modelselection for some cases

[Saurabh7]

No description provided.

[Saurabh7]

@karlnapf
Things should be a bit better now with the fixes here and there. Have a look when you find some time.

The major problem still is using modelselection in web-demos on large dataset and various combination arising.
I think we should disable it when the data is larger than some point, and mention that in the description. That should simplify things and stop a lot of nastiness. Please let me know you thoughts.

[karlnapf]

I agree, there should maybe be a timeout function?
Another alternative could be to provide an interface to cross-validation where people can evaluate the say accuracy/MSE for the parameters given and then do the model-selection themselves.
But for small datasets, its still very nice to have this.

Any ideas how to restrict larger datasets?

[Saurabh7]

hmm I could set up client side timeouts but they won't help afaik and we have no idea about server stuff this is going to run on.

xval can be done but in similar way, exposing potential big computations through the interface will lead to problems, no?

about restricting datasets, maybe just check num of samples (say < 100 for demo purposes) ??

[karlnapf]

yes you are right, exposing potantial big computation through this interface is critical, but i think we can restrict things in exacly that way, only allow for small number of samples. If there are more points, an error pops up saying "Model selection only supported for less than N samples due to computational costs)

[abinashmeher999]

@Saurabh7 @karlnapf are we doing all the computation on the server side and returning the results? Is the large dataset provided by the user? If computation is the issue, I have heard of JApplet that runs on the client side, maybe we can use that. What do you say?

[karlnapf]

That is a good idea in principle, but we currently dont have ressources to think about this. If you wanna come up with something, feel free, but you would be on your own.
For now, restricting demos to toy problems is the way to go.