Research On Partial Label Learning Algorithm Based On Candidate Label Aware And Sparse Reconstruction

Partial Label Learning is a weakly supervised machine learning framework which aims to induce a multi-class classification model from the data where each training example is assigned with a set of candidate labels,among which only one is the true label and cannot accessible during the training phase.The key to learn from such ambiguous labeling information is to disambiguate the candidate label set.Existing partial label learning algorithms only focus on the feature space when calculating instance similarity,and seldom consider using information from candidate label space to construct the similarity relation between instances.And in the disambiguation process,the promotion effect of different labeling confidence levels of the candidate labels and structural information of feature space on disambiguation process is seldom considered.These will lead to insufficient information utilization and unsatisfactory generalization performance of the model.Therefore,this paper studies from the following two aspects:(1)In order to fully exploit the feature space and label space information of training instances,this paper proposes a novel Partial Label learning via Candidate label Aware and Label Propagation(PL-CALP)framework.First,in order to describe the relationship between instances in a more comprehensive way,the algorithm considers the assumption that there should be higher similarity between training samples containing the same candidate labels when constructing the similarity matrix of training instances.Then,the normalized labeling confidence after disambiguation is obtained by iterative label propagation strategy.Finally,the classification model is generalized using multi-output support vector regression.During the model generation process,the algorithm takes full account of the loss of each partial label example to improve the prediction accuracy of the model.(2)In order to fully exploit the latent information of the feature space and different labeling confidence levels of the candidate labels to promote disambiguation process,this paper proposes a novel approach for Partial Label Learning by Sparse Manifold Disambiguation(PL-SMD),which utilizes the structural information of the feature space to facilitate the label disambiguation process.First,this paper describes the underlying structure of feature space through sparse reconstruction of training samples,and the underlying structure information is incorporated into the label disambiguation process based on the manifold assumption.Then,the different confidence levels of candidate labels are formalized into potential label distribution,and present a method which performs label disambiguation and predictive model training simultaneously.Finally,The labeling confidence matrix and classification model are optimized by alternating iteration.In this paper,extensive experiments on five synthetic UCI datasets and six real datasets clearly verify that the two proposed algorithms achieve better classification performance than multiple existing partial label learning algorithms.