This repository is the official implementation of "Guoqiang Wu and Jun Zhu. Multi-label classification: do Hamming loss and subset accuracy really conflict with each other?" accepted in NeurIPS 2020.

The source code is written by Matlab

• ./Datasets -- the datasets downloaded from the websites http:https://mulan.sourceforge.net/datasets-mlc.html and http:https://palm.seu.edu.cn/zhangml/
• ./measures -- the measures for MLC including Hamming Loss, Subset Accuracy and Ranking Loss
• ./Results -- store the experimental results
• ./CrossValidation.m -- used to create cross-validation data
• ./train_hinge_hamming_SVRG_BB.m -- utilize SVRG-BB to train the optimizing hamming loss directly with its surrogate loss (i.e. A^h) where the base loss function is hinge loss
• ./train_hinge_subset_SVRG_BB.m -- utilize SVRG-BB to train the optimizing subset loss directly with its surrogate loss (i.e. A^s) where the base loss function is hinge loss
• ./Predict.m -- predict the model
• ./Evaluation_Metrics.m -- evaluate the model on measures including Hamming Loss, Subset Accuracy and Ranking Loss
• run_linear_hamming.m -- run the code to evaluate A^h
• run_linear_subset.m -- run the code to evaluate A^s

Run the run_linear_hamming.m and run_linear_subset.m in MATLAB, and it will run as its default parameters on sample datasets.