Sign of the IGCI score

[wpzdm]

Hi Diviyan,

I think the sign of the IGCI score is wrong.

In the UAI 2010 paper, on page 5, at the beginning of sec 3.5, Cx2y = S(Py) - S(Px) after preprocessing. And on page 3, after Postulate 2, the paper says negative value of Cx2y indicates X cause Y.

Thus, in method predict_proba of class IGCI, line 105, the entropy estimator should be eval_entropy(x) - eval_entropy(y) (replace x and y). Now, positve return value indicates X cause Y.

Running the code on CEP dataset, the result also agrees with the above.

Best, Abel

[wpzdm]

And it is similar for integral_approx_estimator.

[diviyank]

Hi Abel, This sign is a choice made by design of the package. All pairwise algorithms output a float value between -1 and 1. 1 indicates the causal direction (1st variable causes the 2nd) whereas -1 indicates anticausal direction. The float value is often used to specify the confidence of the algorithm in its prediction.

I do understand your confusion however. I should add a warning/note in the docs.

Thanks for the feedback ! Best, Diviyan

[wpzdm]

Hi Diviyan,

Thanks for the reply!

Maybe I did not make myself clear enough. I was aware of sign of return value of pairwise algorithms in this package from your documentation. Actually, I meant that the return value of IGCI method contradicts the above. That is, given the current implementation, positive return value means anticausal direction (2nd variable causes the 1st).

Best, Abel

[diviyank]

Whoops, you're right, I'll fix that asap

[diviyank]

It should be done, I'll be closing this issue. Don't hesitate to reopen it if the issue remains. Thanks again for the feedback !