Partial Label Learning Algorithm Based On Label Correlation And Partial Label Dimensionality Reduction With Regularization

Learning with weakly supervised information has now become a popular research direction of machine learning.Partial label learning has become an important weakly supervised learning framework,which is successfully applied to a variety of practical problems.Different from the traditional multi-class classification problem,the labels associated with partial label instances are not unique and accurate.Partial label learning aims to learn a multi-class classifier,where each of training examples corresponds to a set of candidate labels among which only one is correct.Because the information in the label space is fuzzy and ambiguous and the ground-true label is invisible to the learning algorithm,the partial label learning is much more complex than the multi-class classification,and it is also more difficult to analyze.The algorithm of partial label learning aims to improve the classification accuracy.The existing methods can be divided into three parts:the algorithm based on disambiguation,the non-disambiguation method and the partial label dimensionality reduction method.The method based on disambiguation strategy extracts the ground-true label from the set of candidate labels and eliminates the ambiguity of the candidate label sets.It can directly deal with the framework of partial label label.The method based on non-disambiguation strategy learns multi-class classification by the one-vs-rest or one-vs-one decomposition strategy.The method of partial label dimensionality reduction combines the disambiguation process and the dimensionality reduction process to solve partial label learning framework,which is conducive to dealing with sparse data and can obtain better feature representations.This article focuses on the two aspects of the disambiguation method and the partial label dimensionality reduction method.The main work includes the following three points:1.The method based on disambiguation is to use the information from the label space or feature space to eliminate ambiguity.Existing methods focus on label space information which is the difference between candidate labels and non-candidate labels,but it is ignored that the label correlation can be learned from the lablel space.To this end,this paper begins with a research of the label correlation in the partial label learning,followed by a unified framework that integrates the label correlation,the adaptive graph,and the semantic difference maximization criterion.Specifically,the label correlation is calculated from the candidate label sets and is utilized to obtain the similarity of each pair of instances in the label space.After that,the labeling confidence for each instance is updated by the smoothness assumption that two instances should be similar outputs in the label space if they are close in the feature space.At last,an effective optimization program is utilized to solve the unified framework.Extensive experiments on artificial and real-world partial label data sets indicate the effectiveness of our proposed method.2.The identification-based strategy is more effective in the current partial label learning algorithm.In the recent years,the identification-based strategy is to get information from the feature space,and eliminate the ambiguity of the candidate label set by this information.Especially,the influence of outlier points can be overcomed by the method of adaptive graph guided disambiguation.However,a linear classifier is be choosed to forecast unseen instances for this algorithm,and the information of label space is not uesd.So in this paper,we introduce the maximum margin method to the method of adaptive graph guided disambiguation to enhance the performance of the classifier and introduce the difference between the ground-true label and the residual label in the label space.The experimental results show that this algorithm is better than the comparison algorithm.3.Partial label dimensionality reduction is a method of dimensionality reduction analysis on partial label data sets.The existing method of partial label dimensionality reduction adopts the linear discriminant analysis method(LDA)and the disambiguation method based on nearest neighbor voting to alternately learn the final dimensionality reduction result.However,the weakly-supervised information in the label space is imprecise and not unique,linear discriminant analysis(LDA)dimensionality reduction will be affected by false positive labels.Therefore,this article adds the manifold regularization term to the LDA method.In this way,the local manifold structure is maintained and the influence of false positive labels is reduced.Experimental results show that this method can improve the classification performance of the original algorithm.