How to transform the output of SAM

[ConnLiu]

Oh my I've just noticed your problem !

The fact is that SAM returns a confidence score for each edge of the graph ( even if it's close to 0). You would be able to see it with the adjacency matrix (networkx.adjacency_matrix(output).todense() i think). Therefore, it plots each edge as it is present in the graph (even if the associated weight is low.)

You must first proceed to threshold the confidence level of the graph before trying to plot it. From experience, thresholding to .6 or .7 leads to obtaining the best results.

Best, Diviyan

Hi, @Diviyan-Kalainathan Your algorithm is really awesome, and I am wondering about how you transform the output of SAM( the confidence score matrix) to a causal graph which is DAG. I have tried thresholding to .6 or .7, but it returns a bad result. As the picture shows below, there are only several edges but a loop.

Originally posted by @ConnLiu in https://github.com/FenTechSolutions/CausalDiscoveryToolbox/issues/65#issuecomment-811597438

[diviyank]

Hello @ConnLiu, Thank you for the kind words. If having a high threshold and still not obtaining a DAG, it means that the algorithm is not able to distinguish betwen the cause and the effect between these two variables; but thresholding so high and still obtaining no DAG might be strange:

• Is the dagloss parameter set to True ?
• If yes, can you report the values of the parameters starting with dag* ?
• Is there duplicated variables in the dataset ? (In the figure you've shown me, there seems to be identically named variables) - The package does not yet handle this type of redundancies, but I should work on it soon !

Best regards, Diviyan

[ConnLiu]

Thanks for you detailed replies.

• the dagloss is True
• Do you mean the parameters of SAM? All is default value except training and testing epochs. And here is the value:

lr:0.01 dlr:0.001 mixed_data:False lambda1:10 lambda2:0.001 nh:10 dnh:10 train:2500 test:1000 batch_size:-1 dagstart:0.5 dagloss:True dagpenalization:0 dagpenalization_increase:0.01 losstype:fgan functionalComplexity:l2_norm sampletype:sigmoidproba linear:False numberHiddenLayersG:2 numberHiddenLayersD:2 njobs:10 gpus:0 verbose:False nruns:2

• No duplicated variables. The names(AST and ALT) are similar but not the value.

Best regards, Conn

[ConnLiu]

Oh my I've just noticed my problem! It is because the confusing figure, in which it look likes a cycle. But actually when I check it by codes[nx.find_cycle(ugraph)], it do not show a cycle.

Really thanks for your help, and sorry about my oversight!