search

jarrycyx / UNN

Causal Neural Nerwork
MIT License
79 stars 9 forks source link

Is it possible that the output causal graph is not a DAG? #2

Closed zmlin1998 closed 11 months ago

zmlin1998 commented 1 year ago

May I sask if it is possible that the output is not a DAG? If that is the case, how can we handle it or can we consider it as a causal graph? Thanks!

jarrycyx commented 1 year ago

Really sorry about the confusion about DAG in the original paper.

Our idea is that, Granger causal graphs in time-series are naturally DAG, since causality does not flow backwards in time dimension. The groundtruth summary graphs (and the discovered CPGs) are not DAGs because we omit the time dimension.

Anyway, thanks for your question.

jarrycyx commented 1 year ago

This point of view is consistent with the works of PCMCI and Neural Granger Causality. However, it does remain a controversial problem if the discovered summary graphs count as causal graphs.

We reopen this issue for possible further discussion.