chschroeder
commented
8 months ago Thanks for the PR!
As discussed, this corrects a deviation of the current implementation from the original paper.
A short summary of our e-mail correspondence so that others can follow:
- Humza Zaki, Vijay Manchiraju, and Dr. Michael Bloodgood found an issue in the implementation the of Predicted Change in F-Measure stopping criterion, where it deviates from the logic of the original paper. Contrary to the paper, the implementation also considered agreement over the negative class.
- Since Dr. Bloodgood one of the authors of this strategy, nobody else could give a more accurate assessment here.
- They investigated the behavior of the observed implementation compared to the original implementation and noted a singificant difference.
- Moreover, they already proposed a fix. Very appreciated :).
- Since they already invested considerable time in finding and evaluating the issue, I offered to do the documentation work (docstrings, changelog, and docs).
I will add the documentation parts and then accept the PR.
zakih2
Pull Request
Description
Updated PCF (base.py) to work as specified in the published implementation.