1,000 samples of 10 variables from a linear-Gaussian model, standardized and rounded to six decimals, seed 4, the columns in an order drawn from the seed. causal-learn's graphs were recorded with causal-learn 0.1.4.8 on the same data.
Correct
Missing
Extra
Wrong direction
PC uses the Fisher-z test at alpha 0.05. GES uses the BIC score with the same penalty and parent limit in both packages. A wrong direction includes an edge that is directed in one graph and undirected in the other.
True graph
CPDAG
The truth is drawn as its CPDAG, the graph the data can identify. Edges without arrowheads are undirected.
Andrey
PC estimate
SHD
3
Skeleton F1
0.95
Arrow F1
0.71
8 correct, 1 missing, 0 extra, 2 wrong direction.
Differences
Missing: the truth has sunlight --> growth
Wrong direction: nitrogen --- growth here, nitrogen --> growth in the truth
Wrong direction: growth --> moisture here, moisture --> growth in the truth
causal-learn
PC estimate
SHD
3
Skeleton F1
0.95
Arrow F1
0.71
8 correct, 1 missing, 0 extra, 2 wrong direction.
Differences
Missing: the truth has sunlight --> growth
Wrong direction: pests --- yield here, pests --> yield in the truth
Wrong direction: yield --> growth here, growth --> yield in the truth
True graph
CPDAG
The truth is drawn as its CPDAG, the graph the data can identify. Edges without arrowheads are undirected.
SHD counts missing, extra, and wrong-direction edges; lower SHD is better. Skeleton F1 scores which pairs are joined and arrow F1 scores arrowheads; higher F1 is better.