Dear colleagues,
We have built a DBN with (1 to t) slices that aims to predict the state of a TERMINAL NODE which represents a forecast at (t+1).
I would like to highlight that the TERMINAL node is of different nature to the PLATE node in our particular case.
If we want to also forecast at times t+2, t+3, ... using TERMINAL NODES, what is the best strategy?
Thank you very much in advance for your help.
Regards,
Oscar PC.
Prediction with multiple TERMINAL nodes
Prediction with multiple TERMINAL nodes
Last edited by oscarpc on Thu Nov 15, 2012 12:04 am, edited 1 time in total.
-
marek [BayesFusion]
- Site Admin
- Posts: 449
- Joined: Tue Dec 11, 2007 4:24 pm
Re: Prediction with multiple TERMINAL nodes
Oscar,
I'm not sure I understand you 100% but here is my shot at your question. I am wondering why you are using the terminal conditions pane rather than just defining your target nodes in the main temporal plate window. To get an idea of what exactly happens, I suggest that you unroll the network and examine it manually. You will see that all nodes in the temporal plate are repeated for each of the time steps. You can define your target node there and just examine it's posterior probability distribution over the time steps that you are interested in. This seems to be the most straightforward way of approaching this. If I misunderstood your intentions, please let me know and, best of all, send me your network -- I will be glad to advise what to do best.
Cheers,
Marek
I'm not sure I understand you 100% but here is my shot at your question. I am wondering why you are using the terminal conditions pane rather than just defining your target nodes in the main temporal plate window. To get an idea of what exactly happens, I suggest that you unroll the network and examine it manually. You will see that all nodes in the temporal plate are repeated for each of the time steps. You can define your target node there and just examine it's posterior probability distribution over the time steps that you are interested in. This seems to be the most straightforward way of approaching this. If I misunderstood your intentions, please let me know and, best of all, send me your network -- I will be glad to advise what to do best.
Cheers,
Marek
Re: Prediction with multiple TERMINAL nodes
Hi Marek,
Thanks for your prompt reply.
>> I am wondering why you are using the terminal conditions pane rather than just defining your target nodes in the main temporal plate window.
My target node is different to the plate nodes. To put a very simple example, imagine our plate nodes are "good student/bad student" given by a set of observations "exam's results" for the last 12 months. Nonetheless, I would like my target node to be "drop/continue in college next month". That is the definition of the target node is different to that one of the plate node. In addition, I can't use the target node "drop/continue" as plate node, since the state will be always "continue in college next month" for last 12 months. This is the what I meant in my previous post. I hope it is a bit clearer.
In addition,I would like to have a set of terminal nodes, where I could predict not only "drop/continue in college next month", but "drop/continue in college in 2 months", "3 months", ... Therefore, I had thought of putting several terminal nodes in order to have a prediction for each of them in time.
Hope my explanation and network make sense.
Thank you again for your help Marek.
Kind regards,
Oscar PC
Thanks for your prompt reply.
>> I am wondering why you are using the terminal conditions pane rather than just defining your target nodes in the main temporal plate window.
My target node is different to the plate nodes. To put a very simple example, imagine our plate nodes are "good student/bad student" given by a set of observations "exam's results" for the last 12 months. Nonetheless, I would like my target node to be "drop/continue in college next month". That is the definition of the target node is different to that one of the plate node. In addition, I can't use the target node "drop/continue" as plate node, since the state will be always "continue in college next month" for last 12 months. This is the what I meant in my previous post. I hope it is a bit clearer.
In addition,I would like to have a set of terminal nodes, where I could predict not only "drop/continue in college next month", but "drop/continue in college in 2 months", "3 months", ... Therefore, I had thought of putting several terminal nodes in order to have a prediction for each of them in time.
Hope my explanation and network make sense.
Thank you again for your help Marek.
Kind regards,
Oscar PC
-
marek [BayesFusion]
- Site Admin
- Posts: 449
- Joined: Tue Dec 11, 2007 4:24 pm
Re: Prediction with multiple TERMINAL nodes
Oscar,
Thanks for your explanation. I'm afraid I still don't understand fully why the target node cannot be part of the main temporal plate. The meaning of a variable is something that you assign yourself as the model builder. If the node "drop/continue in college next month" has two states TRUE/FALSE, the model will calculate the probability of dropping/continuing at each step. True that the probability of dropping may be low in the first months for students who are serious/good. Terminal conditions are variables influenced only by the last time step of the model, which in your case does not seem to be the case. It seems to me a classic DBN model. I suspect, however, that I am still misunderstanding something obvious in your problem -- after all you feel discomfort in following the classical path. If you want to discuss it in detail and share the model with me, you can contact me by Email (marek@sis.pitt.edu).
Cheers,
Marek
Thanks for your explanation. I'm afraid I still don't understand fully why the target node cannot be part of the main temporal plate. The meaning of a variable is something that you assign yourself as the model builder. If the node "drop/continue in college next month" has two states TRUE/FALSE, the model will calculate the probability of dropping/continuing at each step. True that the probability of dropping may be low in the first months for students who are serious/good. Terminal conditions are variables influenced only by the last time step of the model, which in your case does not seem to be the case. It seems to me a classic DBN model. I suspect, however, that I am still misunderstanding something obvious in your problem -- after all you feel discomfort in following the classical path. If you want to discuss it in detail and share the model with me, you can contact me by Email (marek@sis.pitt.edu).
Cheers,
Marek