Enumerable Label not supported.

[ChewyMoon]

Any label that implements the IEnumerable will throw exceptions saying that it needs to have the EnumerableFeature attribute, when I need a Label, not a Feature.

[sethjuarez]

Great question. In this case the machine learning problem will not accept a multi-valued output (hence no support for en EnumerableLabel). What are you trying to accomplish? Maybe I've made a mistake somewhere.

[ChewyMoon]

I do think at least a vector output should be supported.

[normanhh3]

I don't think the intent is to make any arbitrary graph into a potentially significantly sparse array of vectors. Or am I missing something too? On Jun 11, 2015 10:24 AM, "Justin" notifications@github.com wrote:

I do think at least a vector output should be supported.

— Reply to this email directly or view it on GitHub https://github.com/sethjuarez/numl/issues/28#issuecomment-111153393.

[sethjuarez]

@ChewyMoon I'm interested in a use case. I might be missing something (maybe it's something we should support).

[bdschrisk]

Just saw this, a use case would be learning sequences. Such as in one to many or many to many recurrent neural nets.

[sethjuarez]

Sounds good. A couple of things moving forward:

• We need to come up with a way for the learner to select the appropriate IGenerator given the label (for example, the only true multi-class learner implemented so far is the DT)
• Any suggestions for a sequence learning algorithm?

[bdschrisk]

I'm nearly finished with SVM, scoring, multi-class learning (it handles unlimited classes, with support for using any generator and is parallelised by n classes), recommendation (collaborative filtering) and started with recurrent neural nets. It should be fairly straight forward to implement sequence learning as all you are doing is changing the predicted output to a vector and the vector size would be set from either a labelenumerable attribute or the size of the output layer in a neural net.

[sethjuarez]

OK - we should meet about the code. Some of the previous stuff we've merged seems to duplicate some of the other math structures available. Maybe we should have a face-to-face to make sure I am understanding properly. I also want to move to a unified data structure for all models (some graph type data structures). What do you think? Perhaps we should move the convo to gitter:

[bdschrisk]

Reopening issue.

[bdschrisk]

This is partially implemented, see ISequenceModel and Neural Networks.

We need to discuss for other algorithms...