hour discrimination training set

[gngdb]

Started this method of the data assembler to implement a multiclass training set. Haven't actually finished it yet. It does not work. Y matrix needs to be constructed from the hour indexes. Won't fix this at the moment as am working on the test-train discrimination training set first.

[scottclowe]

What are you doing for test/ train dataset similarity measure?

Could it not be simply done with variance feature selection? We want anova test to say test and training are not significantly different?

Sent from my HTC

----- Reply message -----
From: "Gavin Gray" [email protected]
To: "Neuroglycerin/hail-seizure" [email protected]
Subject: [hail-seizure] hour discrimination training set (#19)
Date: Tue, Nov 11, 2014 13:22

Started this method of the data assembler to implement a multiclass training set. Haven't actually finished it yet. It does not work. Y matrix needs to be constructed from the hour indexes. Won't fix this at the moment as am working on the test-train discrimination training set first.

---

Reply to this email directly or view it on GitHub:
#19

[gngdb]

I'm using the classifier itself to decide whether the test/train are similar. Seems like if the training and test are not sepearable to the SVC but preictal/interictal are separable it is more likely to be able to generalise better.

ANOVA test wouldn't reflect the SVM model. Is ANOVA a multivariate model?

Comparing the first and second moments of a multivariate Gaussian would probably be fine. Could just use the KL divergence.

[gngdb]

Although, a classifier could still work perfectly well if preictal/interictal are separable and training/test are separable as well. Think I'll go for the multivariate Gaussian for now.

[fmaguire]

MANOVA is what you are looking for.

[scottclowe]

Yes

On Tue, Nov 11, 2014 at 9:55 PM, Finlay Maguire [email protected]
wrote:

MANOVA http://en.wikipedia.org/wiki/Multivariate_analysis_of_variance
is what you are looking for I think

—
Reply to this email directly or view it on GitHub
#19 (comment)
.