Research On Partial Label Learning Algorithm With Application To Image Classification

Traditional supervised learning requires explicit supervision information,whereas the labeling of supervision information requires lots of manpower and material resources,and even some labeling requires specialized talents.Weak supervised learning does not need strong supervised information,and the framework of weak supervised learning is more consistent with real life scenarios.Therefore,the framework of weak supervised learning has attracted extensive attention in the field of machine learning.Partial label learning is one of the weak supervised learning.In the partial label learning,each instance has a candidate set with multiple labels,however,only one label is the ground-truth.There are three major difficulties in partial label learning.The first is that the labeling information corresponding to the training instance is a label set,rather than explicitly specifying the corresponding ground-truth label.The second is the similarity of labels in the candidate label sets.The third is that the connection between the instance and the label is underutilized.Based on the above analysis,two algorithms based on partial label learning are proposed in this paper.To make full use of the similarity between the instance space and the label space,this paper proposes a partial label learning algorithm via low rank and label propagation to image classification.There are three main innovations of the algorithm.Firstly,it makes full use of the connection between the instance space and label space based on the manifold assumption.The main idea of the manifold assumption is that if the instances having the similarity in the instance space should share the same similarity in the label space.So we can link the instance space with the label space based on the monifold assumption.Secondly,constructing the similarity between the instances based on the low-rank representation matrix.Compared with the similarity construction method based on distance measurement,the low rank representation method has better effect on the high-dimensional data and can construct the global space.In order to reduce the impact of instance imbalance on the low rank representation matrix,sparse constraints are added when constructing the low rank representation matrix.Finally,rather than use the highest probability value as the ground-truth label,the partial label learning is transformed into a multiple output regression problem.Because there may be labels in the matrix with no obvious disambiguation effect.Therefor,in order to reduce the influence of the pseudo-positive instances on the prediction model,it is not used the maximum probability as the ground-truth label.A large number of experiments have proved that using low rank representation matrix to represent the correlation between instances and extending this correlation to the label space has a great effect on classification performance.To highlight the ground-truth label,this paper propose the maximum label confidence for partial label learning to image classification algorithm.There are two main innovations of the algorithm.Firstly,it combines the loss term with a infinite norm of each instance's confidence,for highlighting the ground?truth label in the candidate label sets.There is a big drawback which lies in the average-based strategy there may be some undistinguishable instance after disambiguation.This algorithm improves the disadvantage of average-based strategy and highlights the ground-truth label.Secondly,it makes full use of the similarity between the instances based on the graph laplacian constraint.The goal of the constraint means the similar instances should have the similar output value.And the aim of adding a regulation is to make the highlighted label close to the ground-truth label.Experimental results show that this algorithm has good performance in classification problems.