A Research Of Multi-label Metric Learning Algorithm

Multi-label learning(MLL)is an important research direction in data mining and machine learning.It has been extensively studied in many real-world applications such as text categorization,image annotation,and gene functional analysis.In contrast with traditional single-label problems,samples in multi-label learning have multiple semantic meanings.The output space grows exponentially with the increase of the number of labels,and the traditional supervised classification framework that only considers explicit or single label no longer works.Theoretically,multi-label learning has two fundamental parts: explore label correlations and build the multi-mapping function.However,label correlation and high-dimensionality of multilabel data generate a new problem,which increases the learning difficulty and reduces the prediction power.This refers to the inconsistency problem of multi-label learning.Thus,constructing a robust and compact model to eliminate the inconsistency of multi-label learning has great theoretical research significance,and also has strong application value.The main contributions are as follows:To deal with the inconsistency problem of multi-label learning,this thesis integrates supervised metric learning into multi-label learning problems,and to build a compact relationship between the feature space and the label space.To this end,a supervised metric learning model named with MLMLI is first proposed,which attempts to learn a similarity metric for multi-label data.The basic idea is to incorporate label similarity as weak supervision to assign higher similarity to the pairs of instances with more similar labels.In the second study,a unified model of metric learning and the regression named with MMML is introduced.The metric matrix characterizes the weights and correlations of both the projected features and the labels,based on this,the regression model can be more compact.In addition,the metric learned by MMML has a closed-form solution and thus enjoys high efficiency.In the third study,SeML,a semi-supervised MLL algorithm based on local regression is proposed.Considering the limitation of the regression model for multi-labeled data,for each data sample,a localized regression function is learned based on the local smooth hypothesis.Furthermore,the local regression model propagates to unlabelled data.Finally,this paper demonstrates the effectiveness and superiority of the above algorithms in their respective scenarios through a large number of experiments.