Multi-label sampling based on local label imbalance

• The local imbalance is more crucial than the global one in multi-label data. based measure assesses hardness of MLSOL and MLUL tackle class issue via imbalance. Suitable application situations our two methods are identified, respectively. Class an inherent characteristic data that hinders most learning methods. One efficient flexible strategy to deal with this problem employ sampling techniques before training a model. Although existing approaches alleviate datasets, it actually level within neighbourhood minority examples plays key role performance degradation. To address issue, we propose novel assess label as well approaches, namely Multi-Label Synthetic Oversampling on Local (MLSOL) Undersampling (MLUL). By considering all informative labels, creates diverse better labeled synthetic instances for difficult examples, while eliminates harmful their region. Experimental results 13 datasets demonstrate effectiveness proposed variety evaluation metrics, particularly case ensemble classifiers trained repeated samples original