Research On Weakly Supervised Label Distribution Learning Metho

Label distribution learning is a novel machine learning paradigm for learning mappings from specific instances to label distributions.Unlike single-label and multi-label learning,it assigns a specific description degree to each label,thus preserving more semantic information and alleviating the label ambiguity problem.Existing label distribution learning methods are designed based on data with complete supervised information,ignoring the problem that annotation information of label is often incomplete in real-world tasks.The annotation information required for label distribution learning is the description degrees associated with a set of labels for each particular instance,and complete annotation of all data would result in significant labor and time costs.In addition,due to label ambiguity,it is difficult for annotators to give the exact description degree of some labels for a given instance.Therefore,in the face of the problem of incomplete supervised information that may result from the above reasons,it is of urgent interest to design corresponding weakly supervised label distribution learning algorithms.In this paper,we study the existing weakly supervised label distribution learning algorithms in depth,and propose two weakly supervised label distribution learning algorithms to improve the related work from the perspectives of sample similarity,label correlation and label ranking relation.Meanwhile,a facial expression recognition system is designed based on the two new algorithms for better integration with practical applications.First,this paper proposes a weakly supervised label distribution learning method based on global sample correlation.In weakly supervised label distribution learning,most of the existing works consider the sample correlation from a local perspective and use the k NN method to mine the local correlation of the labels.The sample correlations obtained by this approach largely make the performance of weakly supervised labeled distribution learning degrade as the hyperparameter k changes,thereby,leading to poor algorithm robustness.Therefore,in this paper,we design an algorithm that mines sample correlation from a global perspective to solve the weakly supervised label distribution learning task.Experimental results demonstrate the effectiveness of the proposed algorithm.Second,this paper proposes a weakly supervised label distribution learning method based on local label ranking relation.Existing weakly supervised label distribution learning methods focus on recovering and predicting the numerical magnitude of label distribution values,while ignoring the relative order relationships among these labels.Based on this,we propose a weakly supervised label distribution learning method based on local ranking relation,which can improve the learning and prediction performance of weakly supervised label distribution learning while preserving the relative ranking relation implied in existing label distributions.Besides,we also note that existing weakly supervised label distribution learning methods mostly use low-rank constraints when considering label correlation.However,we found through extensive experiments that this method causes the recovered and predicted label distribution matrices to converge to an identical value(i.e.,the mean value),which results in the absence of relative ranking relation between labels and the increase of label ambiguity.Finally,we design the model to consider not only those label distribution values that are not yet missing,but also focus on the relative error loss of missing values in the regression.Combining the above three points,this paper further improves the performance of weakly supervised label distribution learning while also preserving the relative ranking relation of labels.Finally,we introduce the evaluation metric of label ranking to numerically characterize this ranking relation.Third,this paper implements a weakly supervised label distribution learning based expression recognition system.In facial expression recognition,human facial expressions are often the result of a mixture of multiple underlying emotions,such as happiness,sadness,surprise,anger,disgust and fear.And these basic emotions often express different intensities in a specific expression,thus presenting a wide range of complex emotions.Due to the high cost of labeling time and labor and the limited level of labelers,it is impossible to give accurate and complete facial expression distribution datasets.Using weakly supervised label distribution learning algorithms for facial expression recognition can give multiple emotion description degrees for each object simultaneously and retain facial emotion information more comprehensively.Therefore,we will develop a system based on weakly supervised label distribution learning algorithm for facial expression recognition.