Inference in triangular undirected graphical model

[vladislavs...@gmail.com]

Hi,

I am very new to Graphical Models but have read some introductory texts to understand what it is for and how it works. It would be tremendously helpful whether my particular graphical model can be implemented and perform inference in it. I've got some classification problem which I believe fits well within the Graphical Model paradigm.

Some facts about my data:

- My data is a set of sequences

  There is relatively large number of sequences, say 10000.

- The length of the sequence is variable
  From 3-5 to 10-15.

- Every sequence has a single discrete label [1,...,K]

  The K is relatively small, say 10

- Each element of the sequence is a fixed dimensionality vector (can be reduced to a single discrete variable)

  The number of possible vectors is relatively high, say 1000. The number of possible vectors can be controlled by the process which generates the data,

Some prior knowledge about data:

It is known that one or very few vectors strongly influence the class label C.

The goal is to classify each sequence in one of K classes. To solve this problem, I've come up with triangular undirected graphical model. The model is basically linear chain CRF with an additional hidden variable C, which is connected to each hidden variable. The graph for a sequence of length 3 looks like this:

     C

/     |     \

O - O -  O

|     |      |

F    F    F.

Here the feature vectors are F, the intermediate hidden variables are O and finally the class label variable is C (also hidden).

I am planning to proceed as follows:

(1) training

Set dimensionality and cardinality of hidden variables.

Get training data (F and C)

Train model parameters

(2) testing

Load the trained model parameters

Get testing data (F only)

Perform inference for C (for O as well)

Is this possible to do such learning with libDAI? What is missing or wrong in my plan? Is it possible to perform inference with libDAI in such a graph? What about variable length sequences, is it possible to model this with libDAI?

Thanks