[BUG] ANM + predict_proba method return has wrong

FenTechSolutions / CausalDiscoveryToolbox

Package for causal inference in graphs and in the pairwise settings. Tools for graph structure recovery and dependencies are included.

https://fentechsolutions.github.io/CausalDiscoveryToolbox/html/index.html

MIT License

1.08k stars 198 forks source link

[BUG] ANM + predict_proba method return has wrong #109

Open 945716994 opened 2 years ago

945716994 commented 2 years ago

Encountering a bug while using this package may occur. In order to fix the said bug and improve all users’ experience, it is highly recommended to submit a bug report on the GitHub issue tracker: https://github.com/FenTechSolutions/CausalDiscoveryToolbox/issues

Describe the bug The predict_proba method that in ANM, it said:" 1 if a->b and -1 b->a), so the return statement why is anm_score(b, a) - anm_score(a, b). Why not is anm_score(a, b) - anm_score(b, a).

Please mention

My cdt package version is the newest.
My Python version is 3.8.
My PyTorch package version is 1.8.1.
My GPU is GTX1050, is available.
The full traceback of the raised error if one is raised.
A small code snippet to reproduce the bug if the description is not explicit.

baosws commented 2 years ago

The ANM score for direction a->b (anm_score(a, b)) is the independence test statistic between a and residuals, and lower value = more likely independent (section 3 here) = more likely a->b. In that sense, if a->b is the true direction then anm_score(a, b) should be low and anm_score(b, a) should be high, hence anm_score(b, a) - anm_score(a, b) should be positive. Therefore the implementation is not wrong in this case.

945716994 commented 2 years ago

What about this situation: no edge between X and Y,does the anm_score(x, y) will return 0?

diviyank commented 2 years ago

Hello ! Sorry for the late answer ; yes ; if there is no edge between X and Y the score is theoretically 0. In the real case, the value is close to 0.

Best, Diviyan