On the generator (n=20, T=500, max_lag/tau_max=2), same machine:
|
runtime |
TP |
FP |
FN |
| CausationEntropy, Gaussian oCSE |
61.5 s |
66 |
27 |
11 |
| tigramite PCMCI + ParCorr |
1.9 s |
67 |
30 |
10 |
Comparable accuracy, ~32× slower. Scaling of the Gaussian path: n=5 → 3.2 s, n=10 → 14.9 s, n=20 → 61.5 s, n=20/T=2000 → 125.5 s; n=50/T=1000 did not finish in 6 minutes. The non-parametric estimators are far worse: a single kNN run at n=5, T=500 takes ~130 s, and KDE at n=5, T=500 did not complete in 300 s.
We need runtime/accuracy table against Tigramite.
On the generator (n=20, T=500,
max_lag/tau_max=2), same machine:Comparable accuracy, ~32× slower. Scaling of the Gaussian path: n=5 → 3.2 s, n=10 → 14.9 s, n=20 → 61.5 s, n=20/T=2000 → 125.5 s; n=50/T=1000 did not finish in 6 minutes. The non-parametric estimators are far worse: a single kNN run at n=5, T=500 takes ~130 s, and KDE at n=5, T=500 did not complete in 300 s.
We need runtime/accuracy table against Tigramite.