GDF

[IJCAI'23] Unbiased Risk Estimator to Multi-Labeled Complementary Label Learning (the appendix is shown in the .pdf file)

This code gives the implementation of the paper "Unbiased Risk Estimator to Multi-Labeled Complementary Label Learning".

Requirements

• Python >=3.6
• PyTorch >=1.9

---

main.py

This is main function. After running the code, you should see a text file with the results saved in the same directory. The results will have seven columns: epoch number, training loss, hamming loss of test data, one error of test data, coverage of test data, ranking loss of test data and average precision of test data.

generate.py

This is used to generate complementary labels. After running, you should see a .csv file of complementary labels for a dataset in the vector form. If you have prepared the training data and its complemenatry labels, please ignore it.

Running

python main.py --lo <method name> --dataset <dataset name>

Methods and models

In main.py, specify the method argument to choose one of the 2 methods available:

• GDF: GDF loss function is defined by Equation(12) in the paper
• unbiase: BCE loss is used to MLCLL and derives an unbiased risk estimator, which is defined by Equation(9) in the paper

Specify the dataset argument:

• scene: scene dataset
• yeast: yeast dataset

citation

<code data-enlighter-language="raw" class="EnlighterJSRAW"> @inproceedings{DBLP:conf/ijcai/0003XZ23, author = {Yi Gao and Miao Xu and Min{-}Ling Zhang}, title = {Unbiased Risk Estimator to Multi-Labeled Complementary Label Learning}, booktitle = {Proceedings of the 32nd International Joint Conference on Artificial Intelligence}, pages = {3732--3740}, address = {Macao, China}, year = {2023} }</code>