A Research Of Multi-label Learning Focused On Label Correlation And Label Enhancement

Multi-label learning has been a hot topic in machine learning field.In multi-label learning,label correlation and label enhancement are two of the most important topics.This paper has conducted research focused on those two topics.The content and meaning of this research are as following:Label correlation:Binary relevance(BR)is one of the most popular frameworks in multi-label learning.It constructs a group of binary classifiers,one for each label.BR is a simple and intuitive way to deal with multi-label problem,but fails to utilize label correlations.To deal with this problem,dependent binary relevance(DBR)and other works employ stacking learning paradigm for BR,in which all labels are viewed as additional features.Those works may be suboptimal as each label has its own most related label subset.In this paper,a novel two-layer stacking based approach,which is named SMLS,is induced to exploit proper label correlations.At the first layer,SMLS constructs several binary classifiers in the way of BR.At the second layer,SMLS finds the specific label subset through label selection for each label,and expand them into feature space.The final binary classifiers are constructed based on their corresponding augmented feature space.Comprehensive experiments are conducted on a collection of benchmark data sets.The experimental results validate the competitive performance of our proposed approach.Comparison results with DBR shows that our approach is not only more time efficient but also more robust.Label enhancement:Most multi-label learning approaches treat each label as logical label which indicates whether the corresponding label is relevant or irrelevant to the example.However,for realworld application,logical label cannot express how important the label is to the example.Label Enhancement(LE)approaches aim to transform logical label to numerical label to convey more information.Several LE approaches have been proposed in recent years.Despite their effectiveness in reconstructing numerical labels,the label information was missed in the label enhancement process.To this end,a novel approach named AGLEML is proposed in this paper.First,the topological structure and numerical label are jointly learned,where both feature information and label information are utilized.After that,a predictive model is designed with a few modifications to Multioutput Support Vector Regression(MSVR).Extensive experiments have shown that our approach has achieved superior or competitive performance against the state-of-the-art approaches and could effectively reconstruct label importance from multi-label data.