Research On Label Distribution Learning Combining Dominant Label Information

Label distribution learning is a new machine learning paradigm to solve the problem of label ambiguity,and it has very important applications in dealing with label ambiguity.Unlike multi-label learning,the label space of label distribution learning is no longer a discrete label vector,but the description degree of each label to the sample.It is a more generalized representation.In recent years,label distribution learning has attracted more and more researchers' attention with its wide application background.In the current label distribution learning model,most models use the maximum entropy model to predict the label distribution learning data,and do not utilize the information existing in the label space.At the same time,there is still a lot of room to improve the prediction ability of the model.How to make better use of the information existing in the label space,design a specialized label distribution learning model has become a hot topic in recent years.This paper proposes two special label distribution learning design schemes by introducing ensemble learning,setting dominant labels,and metric learning by focusing on how to make better use of dominant label information and label correlation information in label space.The main work of this paper is as follows:(1)In this paper,a label distribution learning model based on ensemble neural network is proposed.By introducing ensemble learning,neural network and other technologies,the dominant label information of label space and related information between labels are utilized.Firstly,the base learners are divided by the dominant label to ensure the diversity of the base learners.Finally,we use the combined learner to learn the dynamic weights,the learning results of the base learner are weighted to obtain the final label distribution result.(2)In this paper,label distribution learning model based on gemometric mean metric learning is proposed.The dominant label information of the label space is used to adjust the feature space,and a new metric is learned to represent the distance of the samples in the feature space.In this measurement mode,the samples with the same dominant labels are closer,and the samples with different dominant labels are further.Finally,the final label distribution result is predicted by using the k NN algorithm combined with the dominant label information of the labels.