Different Results

[mhsultan-1998]

Hello,

I was told to use this PU algorithm for a research project so I copy-and-pasted the code you wrote into my own Python file and ran it. It has been over a year since you wrote this code and the corresponding article on it so some stuff has changed. Speaking of the article, thank you so much for it. When I ran what you wrote, I got warnings about the label encoder in XGBClassifier and the default evaluation metric in XGBoost having changed. Ignoring those warnings, the results I got didn't match what yours.

Classification results:
f1: 53.54%
roc: 68.28%
recall: 36.56%
precision: 100.00%

When I rewrote the XGBClassifiers as "xgb.XGBClassifier(eval_metric='error')", my results were a little better.

Classification results:
f1: 60.57%
roc: 71.72%
recall: 43.44%
precision: 100.00%

Regardless of what I did with the XGBClassifiers, the baseline results were consistently the same as yours.

Classification results:
f1: 99.57%
roc: 99.57%
recall: 99.15%
precision: 100.00%

Do you know why the final classification results I got are different from yours?

Thank you

[kk-learn]

i got similar results. Using PU -adjusted ML, the classification is a little better than the threshold. i spent so much time on PU methodology which generates all sorts of bugs. I am not going to waste my time on it for my research.

[a-agmon]

The bagging approach is often more stable and gets better results.
Check the bottom part of this notebook for a quick example (under the title "Use bagging and LGBMClassifier")
https://github.com/a-agmon/pu-learn/blob/master/PU_Learning_WBagging.ipynb

@mhsultan-1998 @kk-learn

[kk-learn]