Research On Disambiguation Strategy Based Partial Label Learning Algorithm

In traditional supervised learning,each instance has clear label information.However,with the advent of the era of big data,the time and money costs have increased.Weakly supervised learning has attracted more and more attention.Partial labeling learning can be regarded as a weakly supervised learning framework.This learning framework has a wide range of applications in real society,such as automatic labeling systems,where people from different backgrounds may have different labels,but only one label is the ground-truth.In a report picture,each instance is associated with several names,while the precisely matches between names and faces are not available.Partial label learning aims to learn a multi-classification model from such data set.Existing partial labeling learning only focuses on disambiguation on each instance,but ignores the use of global label information,and most partial label learning algorithm directly use the original feature information,and lack the processing of feature space.Therefore,this paper introduces two partial label learning algorithms.In order to make better use of global label information and to reduce the impact of redundant features,this paper proposes a partial label learning algorithm based on feature subspace representation and label global disambiguation strategy.There are two innovations in this algorithm.On the one hand,we introduced subspace representation in the partial labeling learning.The feature subspace should have three attributes:discriminative,consistency,and compactness.We use the least squares loss to ensure that the feature subspace is discriminative.In order to capture the local manifold structure,we construct projection matrix via graph Laplacian regularization.Besides,we utilize orthogonality constraint to make the high-dimensional features be more compact and avoid redundancy.On the other hand,this model makes full use of global label context to eliminate the impact of pseudo labels in the candidate sets.Global label context can be specifically explained as two principles,the principle of local consistency,similar instances have similar label distributions;and the principle of sparsity,the label confidence matrix should be sparse.Based on the above principles,this model introduces the label confidence matrix.In order to ensure the sparsity of the label confidence matrix,the ?" norm constraint is used to constrain the matrix.In order to ensure the local consistency of the label confidence matrix,the Laplacian constraint is introduced in the label confidence matrix.Extensive experiments show that this method has competitive performance compared with the existing partial label learning algorithms.Due to the existence of pseudo-labels in partial labeling learning,it is difficult for us to learn classification model directly from the candidate label set,so we often disambiguate the candidate label set first.The essential basis for disambiguation is consistency,that is,similar instances should have similar label distributions.Therefore,researchers in partial label learning have proposed various algorithms for exploring instance similarity,including: partial label learning algorithms based on K nearest neighbors and partial label learning algorithms based on minimum reconstruction loss.This paper proposes a partial labeling learning algorithm based on low-rank representation,which innovatively explores the similarity of instances from the perspective of low-rank constraints.Specifically,in the feature space,low-rank constraints are used to learn the self-representation information to fully explore the similarity between instances.In the label space,the self-representation information is used to construct a constraint matrix to make sure similar instances have similar label distributions.Besides the penalty term is used to constrain the label space to ensure the sparsity of the label space.A large number of experiments show that the algorithm has a better learning effect.