Ordinal factor analysis model about infer HOT 6 CLOSED

The issue is that you are using Variable.IsBetween on data where the output is sometimes false. This leads to a weird constraint that requires vYY to not be between two bounds. Do you actually need this for ordinal regression? I would think that each observation in ordinal regression asserts that vYY is between two bounds.

from infer.

By the way, the matchbox recommendation model is similar to ordinal factor analysis.

from infer.

Ah, of course, that makes perfect sense. That was a simple enough fix. I'd like to constrain the weights vectors to be unit length, but it seems that the two more obvious methods for doing so (specifically, taking the inner product with itself or sum of squares and constraining it to be unit 1) don't work. Is this possible under VMP or EP?

from infer.

Thanks much for the prompt reply, by the way. I'll take a look at the matchbox recommendation model.

from infer.

P.S. One more question - is there a straightforward way to constrain an array to be ordered (i.e. for learning the thresholds)? It seems like there should be a way to string together Variable.ConstrainPositive calls to do so but I haven't quite grasped the range syntax well enough yet.

from infer.

It is theoretically possible to have a unit length constraint in VMP or EP. A vector of unit length is equivalent to a point on a hyper-sphere, so the posterior would be a distribution over the hyper-sphere, and various approximate distributions could be used. However, they are not yet implemented in Infer.NET.

You can indeed string together ConstrainPositive calls to constrain an array to be ordered, but you don't need to do this for ordinal regression, since the IsBetween constraints already do it. See the matchbox model.

Since your original question was answered, I'm closing this issue.

from infer.