Algorithm
In the algorithm layer, Causal Learner implements 7 global causal structure learning algorithms, 4 local causal structure learning algorithms, and 15 MB learning algorithms (written in MATLAB). Table 2 lists all of these algorithms. To ensure the correctness of the algorithms implemented in Causal Learner, unless the original implementations of the algorithms are not released, we always try to integrate the original versions rather than
re-implement them. Additionally, we have used the same data to evaluate the algorithms in Causal Learner. Compared with Causal Explorer, the results of Causal Learner are comparable in accuracy and much more efficient.
Table 2: Algorithms included in (indicated by •) and absent from (indicated by ◦) Causal Learner and Causal Explorer.
In the evaluation layer, Causal Learner provides abundant metrics for evaluating causal structure and MB learning algorithms (written in MATLAB). In terms of accuracy, global and local causal structure learning algorithms are evaluated using the same 7 metrics: Ar_F1 (Ar denotes arrow), Ar_precision, Ar_recall, SHD, Miss, Extra, and Reverse. MB learning algorithms are evaluated using 3 metrics: Adj_F1 (Adj denotes adjacent), Adj_precision, and Adj_recall. In terms of efficiency, global causal structure learning algorithms are evaluated using running time, while both local causal structure and MB learning algorithms are evaluated using running time and the number of conditional independence tests.