| Semi-supervised partial label learning combines two difficult learning paradigms,one is semi-supervised learning,the other is partial label learning.Due to the development of Internet technology,the amount of data increases sharply,and it is difficult to obtain accurate data sets.Usually,an example only roughly matches a group of labels,and most data have no labels.Traditional learning algorithms have low generalization performance for datasets that are not accurately labeled,so a new learning framework has been proposed in recent years,namely the semi-supervised partial label learning framework.However,there are still some problems in these solutions.For example,SSPL ignores the consideration that the noise in the candidate tag set of biased labeled samples will pollute the unlabeled samples.The PARM algorithm has the problem of high computational cost when dealing with high-dimensional feature dramas.The contributions of this paper mainly include the following two parts:(1)The application of semi-supervised partial label learning algorithm based on attention mechanism in network image classification is studied.This paper proposes a semi-supervised biased label learning method based on attention mechanism(ASSPL).This method attempts to introduce the attention mechanism into the semi-supervised partial labeling algorithm,using attention to more comprehensively consider the distribution of samples,so as to better spread and classify labels,so as to obtain more accurate classification results.The contributions of the proposed algorithm include: Introducing the attention mechanism into semi-supervised partial label learning,which provides a new idea for learning;ASSPL is a label propagation method based on attention mechanism;Experiments verify the effectiveness of this method in network image recognition.(2)Semi-supervised partial label learning algorithm based on reliable label propagation is studied.This paper introduces a semi-supervised partial label learning algorithm(See PLL)based on reliable label propagation.The existing algorithms do not consider the effect of disambiguation,which is largely affected by the false positive tags of the candidate tag set.In addition to the real tags,the false positive tags in the candidate tag set are essentially noise.In the iterative process,the identified real label may become a false positive label,and the modeling output of the real label will be submerged by the false positive label.Based on this situation,a semi-supervised partial label learning algorithm based on reliable label propagation is proposed,which can make better use of unlabeled data,reduce the impact of label set noise in partial label and improve the prediction accuracy of the algorithm.The contributions of the proposed algorithm include: See PLL propagates the disambiguated reliable label to the unlabeled data set,so that the unlabeled bias can be labeled more accurately;See PLL reduces model parameters to present robust performance on noisy data sets;See PLL provides an optimized solution to adapt to real-world application scenarios.Experimental results on real data sets verify the effectiveness of the algorithm. |
