Subscribe to DSC Newsletter

Dear,

Have some of you already applied successfully neural networks or support vector machines in some real-life industry (bank, insurance, retail/distribution, public services, manufacturing, telecom, pharma...).
Have you already deployed such models for decision-makers and not only as part of model assessment or comparison?

Could you share some experience? industry, product, project...

Thanks in advance for any reply.

Patrice

Tags: data mining, neural network, support vector machines

Views: 849

Replies to This Discussion

Patrice,

We do this on a regular basis, as do many of our software users. For some ideas on what NNs have been used for in real situations please see www.tiberius.biz/casestudies.html

This might also be of interest. A cross sell problem where we used an ensemble of neural nets to come up with a solution that proved as good as any other submissions in a recent data mining competition. Note that using ensembles from a particular modelling techniques and then combining them with ensembles from other techniques will give you the most accurate and robust solution on real life noisy data.

www.tiberius.biz/pakdd07.html
Thanks for your reply, Phil.

Does Tiberius combine ensembles of "weak" NNs?
If yes, does it estimate an optimum number of NNs to combine because I guess it could quickly become a huuuge factory of calculations...

;-)

Patrice
Hi Patrice,

Yes, Tiberius does build ensembles of neural networks. You can make the nn 'weak' by specifying only 2 hidden neurons and not many learning epochs. Each net can be built on a random selection of variables with a random selection of examples. This is a bit like random forests.

The optimum number of nns to combine - well, the more you have the more robust the model will be, so it depends on how much time you have. There is no need to worry about the calculations though, the ensemble model code is spat out in SQL,SAS,SPSS etc.

Cheers,

Phil
Hi Phil,

Could you use the 'weak' NN but instead of random sampling employ adaptive sampling (or reweighting)? The analogy you made re Random Forests reminded me of 'stumps' - adaptive models built from very simple trees.
Hi Patrice,

Is there a reason that Gradient Boost Machines aren't on your list?

Mark
Hi Mark,

My reasons were basically
- not a too long list here (?)
- got some remarks recently specific about SVMs and NNs not used on (Belgian) market
- I don't know Gradient Boost Machines... but I'd like to

Could you share some quick "in a nutshell" intro on these GBMs, please? I would deeply appreciate.

It looks like it uses some boosting principles (and you refer to these in your reply to Phil too). I made my PhD on bagging and boosting so I could probably quickly jump into.


Thanks in advance for your input

Cheers,

Patrice
If your dissertation was on boosting, you should pick right up on GBM (aka "TreeNet" at Salford Systems). The seminal paper is from Jerome Friedman (1999) "Greedy Function Approximation: A Gradient Boosting Machine"

Also, Greg Ridgeway has contributed a "GBM" package in R.
addendum: attaching an interesting paper from Jerome Friedman comparing SVM and Boosted Trees that you might like.

Also, it gives the year of publication for the Friedman GBM Paper I cite above as 2001 in the Annals of Statistics v. 29 pp. 1189-1232
Attachments:
I like your references very much. Thanks a lot!

I haven't studied these techniques in depth nor applied them in any industry after my PhD.

Same question then: what kind of industries already applied GBM successfully according to you? Salford Systems doesn't seem to indicate a lot on this while they argue that their RandomForests are of major interest in biomedical and pharmaceutical R&D which I can confirm from my side.
Thanks a lot for your nice paper!

RSS

On Data Science Central

© 2019   AnalyticBridge.com is a subsidiary and dedicated channel of Data Science Central LLC   Powered by

Badges  |  Report an Issue  |  Privacy Policy  |  Terms of Service