Research On Partial Label Learning Algorithm From Semi-Supervised Perspective

Partial label learning is an important branch of the weakly supervised learning framework.In recent years,the research enthusiasm has gradually increased,and it has a wide range of application scenarios in real life,such as big data,computer vision,and medical information.In the partial label learning task,an instance in the data set corresponds to a set of candidate labels,and there is one and only one true label corresponding to the instance in the set,and the goal of the partial label learning task is to identify the true label.In this scenario,the traditional supervised learning method can no longer be used,and a unique partial label learning algorithm needs to be designed according to the characteristics of the data set.An intuitive way to solve the problem of partial label learning is to perform disambiguation,that is,to eliminate noisy labels from the candidate label set.According to disambiguation strategies,existing partial label learning algorithms can be divided into average disambiguation strategy,discriminative disambiguation strategy and nondisambiguation strategy.The average disambiguation strategy treats each label in the candidate label set equally,and makes the prediction by averaging output.The discriminative disambiguation strategy treats the true label as a hidden variable and performs disambiguation through iteration with other parameters during the model training process.Non-disambiguation strategies transform the partial label learning problem into something else.Although the above method solves the problem of partial label learning partly,when the data presents high ambiguity,there are still some problems like model performance degradation,and the lack of distinguishing processing of supervisory information.Based on this,this paper proposes the PLSP(partial label learning from semi-supervised perspective)algorithm that transforms the partial label problem into a semi-supervised problem.Combined with the current advanced semi-supervised learning method,a semi-supervised learning framework suitable for partial label learning is proposed.The main work of this paper is as follows:(1)Constructing a partial label learning training set under the semi-supervised framework: first,this paper samples data from the training set in the pre-training stage,and trains the model based on the sampled data and a non-disambiguation loss item.Because the sampled data is small,the pre-training can converge quickly;then we use the pre-trained model to make predictions,and obtain the maximum class confidence for each instance in the training set,and compare it with the threshold,if the maximum class confidence is larger than the threshold,the current instance belongs to labeled training set,otherwise it belongs to unlabeled training set;for each instance in the training set,the class corresponding to the maximum confidence is used as its pseudo-label.At this time,the partial label learning training set consists of two parts: a labeled training set and an unlabeled training set.Now the structure of the partial label learning training set is similar to the semi-supervised learning training set,which facilitates the application of the semi-supervised learning framework.(2)Proposing a semi-supervised learning framework suitable for partial label learning:first,a supervised loss is introduced for the labeled training set;then data augmentation(including strong augmentation and weak augmentation)is applied to the instances in the unlabeled training set,and the consistency regularization is applied to the model output;then the model is divided into two parts: the deep feature extraction function and the fully connected layer,then we use the deep feature extraction function to extract the augmented variant depth features,and construct the covariance matrix,then we sample the semantic direction in the Gaussian distribution,and apply semantic transformation to the augmented variant according to the extracted semantic direction,and the transformed deep features are entered into the model,and consistency regularization is applied to the model output,so the regularization term is rewritten;at the same time,this paper calculates the upper boundary of the consistent regularization term,and replaces the original regularization term with the upper boundary,which effectively improves the calculation efficiency;all of the work mentioned above defines the loss item from a semi-supervised perspective,based on the perspective of partial label learning and in order to give full play to the constraint effect of the candidate label set on the prediction results,a supplementary loss is introduced for the instances in the unlabeled training set,making the model prediction result fall in the set of candidate labels corresponding to the instance as much as possible;last,we apply semantic transformation to instances in the labeled training set,and replace the original deep features with the semantically transformed deep features.Likewise,semantic transformations are applied to the supplementary loss term,enabling complete application of semantic transformations on the objective function.This paper conducts a large number of experiments on different data sets,and verifies that the PLSP algorithm proposed in this paper has better performance and execution efficiency than multiple partial label learning algorithms.