GES Algorithm Fails: parameter graph should be Dag or np.ndarray

[Hennex945]

Description
When running the GES algorithm, a type error is raised, despite the input being correctly formatted as a numpy.ndarray. The issue appears to stem from an internal type mismatch during the processing of the graph object within the GES, BOSS, and GRASP implementation.

The issue persists with both real-world and minimal synthetic datasets, indicating a potential problem in the library's handling of the GeneralGraph type.

Analysis

• The error originates from the dag2cpdag function in the DAG2CPDAG.py file.
• The GES implementation in GES.py passes a GeneralGraph object to dag2cpdag, but this function only accepts inputs of type Dag or np.ndarray.

Steps to Reproduce
The issue can be reproduced with the following minimal example:

from causallearn.search.ScoreBased.GES import ges
import numpy as np

data = np.array([[1, 2, 3],
                 [4, 5, 6],
                 [7, 8, 9]])

record = ges(data)  # Expected to work, but fails

Expected Behavior
The GES function should successfully execute and return a Record object containing the causal structure inferred from the input data.

Proposed Solutions

1. Modify dag2cpdag: Update the function to handle GeneralGraph objects, converting them to Dag if necessary.
2. Adjust GES Implementation: Convert the GeneralGraph object to Dag before calling dag2cpdag.

Environment

• Python Version: 3.8.10
• CausalLearn : 0.1.3.9

Thank you for looking into this, and please let me know if further details are required!

[kunwuz]

Thanks for reporting this! This seems to be a bug in the latest version (0.1.3.9). We will fix it asap. At the same time, perhaps you could try using 0.1.3.8, which should work.

[kunwuz]

hi @Hennex945 , could you please test the latest code in the repository and let me know if the issue still persists? Thanks!

[Hennex945]

Hi @kunwuz, thank you for the fast reply. I tested the latest code and the type error is fixed. Unfortunately, I receive a follow-up error in the LocalScoreFunction.py file (in my case in the method local_score_BIC_from_cov).

Singular matrix
  File "causallearn\score\LocalScoreFunction.py", line 73, in local_score_BIC_from_cov
    H = np.log(cov[i, i] - yX @ np.linalg.inv(XX) @ yX.T)
  File "causallearn\score\LocalScoreFunctionClass.py", line 44, in score
    self.score_cache[i][hash_key] = self.local_score_fun((self.cov, self.n), i, PAi, self.parameters)
  File "causallearn\utils\GESUtils.py", line 11, in feval
    return parameters[0].score(parameters[2], parameters[3])
  File "causallearn\utils\GESUtils.py", line 238, in insert_changed_score
    score1 = feval([score_func, Data, j, tmp3, parameters])
  File "causallearn\search\ScoreBased\GES.py", line 214, in ges
    insert_changed_score(
  File "test.py", line 42, in <module>
    record = ges(data)
numpy.linalg.LinAlgError: Singular matrix

I will further investigate the problem and open a follow-up issue if necessary.

@edit: I just realized I used a singular matrix and the error is my fault. I tested two different use-cases, and they just worked fine. Thanks again for your quick help!