How can I generate a graph from a data without using cdt.data.AcyclicGraphGenerator

[MAX00008888]

[MAX00008888]

@Diviyan-Kalainathan , sorry to bother you. I don't understand the difference between create_graph_from_data(data, **kwargs) and orient_directed_graph(data, dag, **kwargs)``predict(df_data, graph=None, **kwargs). How can I generate a graph like this waydata, graph = generator.generate().

[MAX00008888]

how to get a graph contains the ground truth of the dataset @Diviyan-Kalainathan

[diviyank]

Hello @MAX00008888,

There are two distinct things :

• cdt.data that generates/provides data with the ground truth
• cdt.causality methods, which goal is to find the ground truth graph that generated the data

Actually, predict regroups multiple functions for causal discovery methods, one that cover each case, and runs for each a different function:

• The user provides only the data (we have to produce a graph from scratch : create_graph_from_data()
• The user provides the data and the skeleton of the graph (undirected graph) : orient_undirected_graph()
• The user provides the data and a (even partially-) directed graph : orient_directed_graph

Hope this helps, Diviyan

[MAX00008888]

Thank you soooo much for your relpy @Diviyan-Kalainathan ,

I'm still a little confused. Here is my code: data= pd.DataFrame(batch.next_observations.cpu().numpy(), columns=cols) model = GES() graph = model.create_graph_from_data(data) Is this get the same kind of graph like data, graph = generator.generate().

Thank again!

[diviyank]

Yes, GES aims at producting the same graph as the one that generated your data (here batch.next_observations.cpu().numpy()) ; it tries to find causal relationships between the columns of your data.

Best, Diviyan

P.S. : be warned however, GES aims to produce the ground truth of your data and therefore the results are not always 100% correct; compared to the AcyclicGraphGenerator (that actually generates data according to the graph), the causal methods do not provide the true answer (it's their goal, but it's the whole point of the research)

[MAX00008888]

Thank you so much ! !!@Diviyan-Kalainathan It may sound silly , but I would like to know if there is anyway to get the true graph of data.

Thanks

[MAX00008888]

Hi Diviyan, @Diviyan-Kalainathan

I want to know if there is a score function like BIC score to get the score of the graph.

Best,

[diviyank]

Hi,

Sadly, the causal methods try to find the true graph of the data (or the markov equivalence class - check out CPDAG), and more often than not, the assumptions made by the used models are violated. Thus the approximate result has not optimal guarantees on the validity of the proposed solution.

We don''t have a score function such as BIC in the toolbox, but that's a good idea ! I will look into it.

Best regards, Diviyan

[MAX00008888]

Thank you so much for your reply!!! So sorry that I forgot to reply you, you can close this question.

Best wishes.

[diviyank]

Hello! You're welcome! I'll just keep this issue open to add the scores in the development roadmap