facebookresearch / mmd Goto Github PK

Make library into a Python pip package (and Improve code organization in the process)

Provide illustrative examples of how to use the MMD library both in README.md and as part of an example code file (preferably a computational notebook, e.g., Jupyter)

RT

Hello,

I was trying to understand the code implementation of your FSE21 paper.
In section 4.2, the maximization objective of LearnRule(L, A) (which I identified as the function generate_conjuncts()) was defined as:

lambda1 * precision + lambda2 * recall + lambda3 * (1 / size(rule))

where size(rule) is no. of predicates in it. However, in the code I see the objective function written as:

# Guide rule size with cfg.max_size parameter
size_score = 1 - self.get_size() / cfg.max_size

q = stats.get_coverage_ratio(self) * cfg.coverage_conjunct \
    + stats.get_tp_ratio(self) * cfg.tpr_conjunct \
    + size_score * cfg.size_conjunct

This makes the interpretability objective some_constant - size(rule), and not some_constant / size(rule).
I understand they try to achieve the same thing for the overall objective (minimizing rule complexity in terms of no. of predicates), but I was wondering whether the choice was made because of numerical stability or for any other reason. Your insights on this will be much appreciated.
Thank you for taking the time to make the code public. It has been very helpful.