Multi-Label Learning Based On Metric Learning And Optimizing The Ranking

Multi-label learning is a natural extension of traditional single-label classification.In this framework,each example's label information is represented by a subset of the label space.Therefore,the study of this learning paradigm has not only theoretical significance,but also extensive application value.At present,there are many excellent research results on various aspects of multi-label learning problems.The research content of this paper mainly focuses on three aspects,including the following works:First of all,existing research on features in multi-label learning is mainly focused on the field of dimensionality reduction.There are only a few research advances in learning better distance metrics and reducing interference from characteristic noise.Based on the information geometry metric learning algorithm,we introduce the cosine similarity function and Jaccard similarity coefficient to measure the similarity of the set of labels corresponding to each pair of samples,making it suitable for multi-label learning.And the squared Frobenius norm of the metric matrix is used as a regular term to ease the occurrence of overfitting.Experiments on the multi-label datasets confirm that the new distance between samples calculated using the metric matrix can improve the performance of the multi-label classification algorithm.Secondly,when the multi-label classification problem is transformed into the label ranking problem,the information in the label space is redundant,that is,some label information cannot improve the performance of the existing algorithm,but increases the computational cost.This paper establishes a different constraint to deal with the above problem,focusing on labels with greater uncertainty,and minimizing the slack terms introduced when the above constraints are not met and the complexity of weight parameters to achieve label ranking.Simulation experiments on some multi-label datasets confirm the effectiveness of the algorithm.Finally,considering the problem of high computational complexity of Rank-SVM during model training,this paper introduces a weighted linear loss to approximate the hinge loss function.The solution is more efficient by modifying the regular terms in the objective function.Compared with the original algorithm,although the performance on the evaluation criterions has slightly decreased,it has an absolute advantage in computational efficiency.