How to observe variables in large size network, using set_evidence() and update_beliefs()

[BayesFusionUser123]

# How to observe variables, set_evidence() on large network and update_beliefs()

I want to observe several variables and update_belifes() and then check the posterior probability of the outcome variable of interest.

I am write code as follows

Code: Select all

```
net.set_evidence(node_name, val)
net.update_beliefs()
```

I tried on one variable first and the code `net.update_beliefs()` is very slow, I checked after hours, it still not print out result.

The net is a parameterized network using EM algorithm. The graph statistics is as follows:

Code: Select all

```
=== Bayesian Network Graph Statistics for SMILE ===
Is Directed Acyclic Graph (DAG)?: True
Number of Nodes (Variables): 1320
Number of Edges (Direct Dependencies): 3752
Max Parent Count (Max In-Degree): 14
Weakly Connected Components: 28 (Inference can be segmented if > 1)
Moral Edges Added (Co-parent marriages): 5311
Approximate Treewidth (of moral graph): 370
```

I have the following questions:
1. If I set_evidence() for one variable, does `net.update_beliefs()` need to calculate probability information for all nodes in the graph?


2. In SMILE wrappers document, I saw the following description when Network.updateBeliefs has problem:

Code: Select all

```
Network.updateBeliefs throws an exception with error code -42 if the temporary data structures required to
complete the inference take too much memory. In such case, or if the inference takes too long, consider taking
advantage of SMILE's relevance reasoning layer. Relevance reasoning runs as a preprocessing step, which can
lessen the complexity of later stages of inference algorithms. Relevance reasoning takes the target node set into
account, therefore, to reduce the workload you should reduce the number of nodes set as targets if possible. Note
that by default all nodes are targets (this is the case when no nodes were marked as such). If your network has
1,000 nodes and you only need the probabilities of 20 nodes, by all means call Network.setTarget on them.
```

I have a outcome variable named "O" in a large network, I want to observe all direct causes and direct effect of this outcome (i.e. all parents and successors of outcome variable "O").

2.1 Could you please write some sample code for to set targets?
2.2 In my case, should I "set target" for the following nodes: 1) outcome variable "O", 2) all parents and all successors of "O"?
2.3 What is the difference between following two cases: 1) observe on the original network (without "set target") and 2) "set target" on the network and then do observation?

Thanks.

[piotr [BayesFusion]]

1. If I set_evidence() for one variable, does `net.update_beliefs()` need to calculate probability information for all nodes in the graph?

Without a target set up before update_beliefs SMILE will perform a full network update.

The set_target should significantly speed up operations within update_beliefs. You should set as targets all the nodes whose values you're interested in - in this case, all the children and parents of node O.

What is the difference between following two cases: 1) observe on the original network (without "set target") and 2) "set target" on the network and then do observation?

These operations are independent. As long as you don't run update_beliefs, it doesn't matter which one you execute first.

Sample code:

Code: Select all

for p_handle in net.get_parents("O"):
    net.set_target(p_handle, True)
    
for c_handle in net.get_children("O"):
    net.set_target(c_handle, True)

[marek [BayesFusion]]

Hi BayesFusionUser123,

You are facing problems with the complexity of your model. While 1,000+ nodes is nothing special, it all depends on the connectivity of the network and also max in-degree, which in your model is 14. Even if your variables are binary, the node and all its 14 parents will be part of a huge clique and the clustering algorithm may start choking. As Piotr suggested, one thing you can do is focus your inference, i.e., designate those nodes that you are interested in as targets. This should significantly reduce your problem. You can try this all in GeNIe first -- you will know whether the model can be updated at all. Another thing that you can do is restructure your model so that it does not have nodes with that many parents. Please consider, for example, the "parent divorcing" technique, described in the manual. If none of these help, you can also use a stochastic sampling algorithm -- it is approximate but gives quite good results. I recommend EPIS-BN, quite likely the best sampling algorithm in existence.
I hope this helps.

Marek