Computational Time and structure learning

The front end.
Post Reply
Gardiola
Posts: 29
Joined: Mon Mar 11, 2013 4:38 pm

Computational Time and structure learning

Post by Gardiola »

Good Morning,

I have 130 000 records and 8 variables.

I launched the PC learning algo for structure learning !

my computer is core i7 ( 4 cores ) and 16 GBram.

It s still running after 15 minutes....

my question is : After how many minutes we can decide to stop Genie and resuggest to Genie a shorter list of variables ?

Thanks

Didier
shooltz[BayesFusion]
Site Admin
Posts: 1417
Joined: Mon Nov 26, 2007 5:51 pm

Re: Computational Time and structure learning

Post by shooltz[BayesFusion] »

Gardiola wrote:my question is : After how many minutes we can decide to stop Genie and resuggest to Genie a shorter list of variables ?
The only criterion in such case is the amount of time you can spend waiting. If you can specify such length of time, I suggest using 'Max time' option in PC parameters; with approach you will avoid cancelling the PC run and will be able to see the structure produced by the algorithm.
Gardiola
Posts: 29
Joined: Mon Mar 11, 2013 4:38 pm

Re: Computational Time and structure learning

Post by Gardiola »

Ok.

I decreased the number of variables as well.

By the way, I used Genie during all this afternoon: I leveraged Genie as a Tool to mine the data as well.

The only comment is that I can made it will be good if the last target value can be kept in order to compare with the new value.

also, I don t understand how the sensitivity analysis work:

1) I defined the target.

2) i click on the icone.

3) the red variables have an high impact on the target ?
shooltz[BayesFusion]
Site Admin
Posts: 1417
Joined: Mon Nov 26, 2007 5:51 pm

Re: Computational Time and structure learning

Post by shooltz[BayesFusion] »

it will be good if the last target value can be kept in order to compare with the new value.
I'm not sure what you mean by last target value.
the red variables have an high impact on the target ?
The color of the node when sensitivity mode is enabled is calculated by finding maximum sensitivity value for node's CPT.

Each entry in the CPT has its own sensitivity (here defined as the value of derivative at x of a function y=P(x); y = posterior of the target, x = value of the specific parameter in the CPT). Since targets can have more than two states and more than one target can be enabled for sensitivity calculation, each CPT entry may have multiple values for each {target , target outcome} pair.

Per-node maximum is then divided by per-network sensitivity maximum so color values are normalized. Effectively, the red variables have highest impact on the target(s).

You can view the actual sensitivity values in node's CPT table. With sensitivity enabled the background color of CPT cells is based on their respective sensitivity. For actual numeric value, hover the mouse cursor over the cell to show the tooltip.

Note that sensitivity depends on the evidence set.
Post Reply