Thank you for the really amazing help, Stephen.
Appreciate all your comments. I will read the doc you shared now and try to see how Kalman filter could fit there. It seems like a good approach, but so did the one using Block HMMs... so, guess will have to do some more reading.
Yes, the degree can (and usually does) changes from week to week.
It could be that C and D describe a state, so it's not C only. I agree it should be called "state variable".
I don't see any emission between two states. It is important to consider previous state when inferring the current...
Numerical value simply indicate to what extent student "belongs" to group C. That is, observing his interactions with other students, how much of his activity is associated with other students in the group C.
Here we are talking about clustering again. There are (at least) to cluster students...
It is a single state. It just that we can say that cluster 1 represents this state. So, when we say (for example) that students most commonly transition from state 1 and state 3 to state 2, we should be able to say that students usually transition from C and B to A and D (from the previous example).
Each student has his/her own trajectory, but the model doesn't depend on student's age or other "properties". Also, it should be a finite number of states. I didn't think about conditions, but each states should be defined with certain values of A, B, C, and D. For example, State 1 is described...
That example would work as well. The difference between the problem you explained and the one I'm trying to solve is the following: I don't have "rain", "sun", "snow". Also, I don't have "1", "2", "3". Instead of those discrete values, I have
day 1 (0.2, 0.3, 0.5, 0.0)
day 2 (0.8, 0.0, 0.0...
Hi Stephen,
I am familiar with graph theory (at least more than HMMs), and the only reason I posted here was to ask whether there is an HMM implementation that allows for describing each observation in a sequence/trajectory with feature vector, instead of a single (discrete or continuous)...
You are right, Stephen.
Each state should be "described" with one or more "clusters". Those probabilities are actually obtained using Mixed Membership Stochastic Blockmodels (MMSB), applied on the graph of learners. I didn't want to go into all the details since I thought that what I said...
Thank you for the great explanation, Stephen.
I agree that state is not clearly defined here and I'm trying to apply HMMs in a situation that is not very well defined. Let's make clear several things. Here I have trajectories for several students. When I specified the model, I basically defined...
Thanks mathman and mfb,
Really sorry for the confusion. Let me try to provide better explanation. The sample from my dataset would look like this:
sequence parent_i week student cluster1 cluster2 cluster3 cluster4 cluster5 cluster6 cluster7
0 0 w2 1111 0.032 0.007...
Hi all,
Not sure if this would be the right place for this question, but I know it bothers me for some time already and would really appreciate any kind of help. I am trying to fit an HMM, but here for every observation in the sequence I have feature vector - probability distribution that given...