The code of Oversampling Multi-Label Data based on Natural Neighbor and Label Correlation: step:
1.pip install skmultilearn or move skmultilearn we provide in your conda env (I suggest moving the package we provide directly,because the original package had many bugs).
skmultilearn provides many classic MLC algorithms, such as BR, CC, RAKEL, and MLkNN, etc. but unfortunately, it is no longer maintained and we have fixed bugs in these algorithms. More details http://scikit.ml/.
2.download MLNONC.rar
You can adjust the sampling rate p from the Main function. The code of MLONC is mainly divided into three parts:
data processing (Each function is easy to understand) Sampling (MLONC main Code) our main method (searching for natural neighbors and sampling) with training(5 * 2 fold cross validation)
main package version:
numpy 1.21.6
pandas 0.24.2
scikit-learn 1.0.2
scikit-multilearn 0.2.0
and python==3.7
We have provided Macro-F for evaluation metrics, and more evaluation metrics can be easily added by yourselves. Of course, we will provide all parameter values after the paper is accepted.
You can follow me, and I will create a Python based multi-label sampling algorithm library in March, including MLROS, MLSMOTE, MLSOL, REMEDIAL, MLTL, etc. And it will update some algorithms that are not provided by skmultilearn, such as ECC and COCOA In addition, due to the popularity of deep learning, I will also upload a summary of deep learning algorithms, including C2AE, MPVAE, CLIF, PACA, DELA.