Using background knowledge makes FCI algorithm slower about causal-learn

[jdramsey]

The issue I think is that checking background knowledge requires an overhead, so if you don't have to do it the algorithms will run faster.

The way it's stored is in pairs <{n1, ..., nm>, {nq...nr}>, where edges are forbidden in reverse. The set lookups should be fast (i.e., if you forbid all of the second from causing any of the first, and you try to add an edge x1->x2, if will look to see if x1 is in the second set and x2 is in the first set and if so will forbid it). If you give separate singleton pairs for each, like <{x1}, {x2}>, it has to go through a lot of sets to figure this out. Do you know how your knowledge is being entered?

[verae98]

Thank you for your reply! I store the forbidden and required edges in self.forbidden_rules_specs and self.required_rules_specs for the BackgroundKnowledge class which I thought only allow singleton pairs, since the type is Set[Tuple[Node, Node]]. That explains why the algorithm is slower then. But how can I change then then to the pairs with multiple values?

[verae98]

Hi, I have added a childclass of this BackgroundKnowledge class called FastBackgroundKnowledge in my code which inherits all the properties of the BackgroundKnowledge, except for the is_forbidden() and is_required() functions to fix my problem.

I have added two dictionaries for which the required and forbidden edges are stored in. For each required edge X1 -> X2 we have a dict_required {"X1" : [X2]}. This way the is_forbidden() function and is_required() function do not have to look through all sets, but has a lookup by node name. This has lowered the FCI function computation time significantly!

There is only one problem, in the file fas.py line 286, type checking has been done by stating type(knowledge) == BackgroundKnowledge I would like to suggest to change this to isinstance(knowledge, BackgroundKnowledge), such that inheritance is possible :)

[kunwuz]

Thanks so much! Your solution looks fantastic : ) If you would like to, could you please create a PR (including a short test example) for this so that your contribution could be better recognized? We can also review and push this if everything goes smoothly. Of course, we greatly appreciate your suggestions whether you would like to PR it or not.