Multi-label Classification And Label Completion Algorithm Based On K-means

Label learning is a framework that assigns appropriate category labels to the unseen examples.Each example in single-label learning has only one label,for example,a digital image is passed to the computer,and the program returns a prediction about the number,but only a single number.There is a certain difference between multi label learning and single label learning.In multi label learning,each example has multiple labels associated with it.For example,if you send a street image to the computer,what the program returns is to determine which objects are in the image.In this thesis,we propose algorithms for multi label classification and how to complete the label when the label is missing.The main contents are as follows:(1)In multi-label learning,correlation not only exists in feature space or label space,but also exists between them.The traditional K-Nearest Neighbor algorithm is a simple and easy to understand multi-classification algorithm,which has a good effect on the correlation.The existing algorithms based on k-nearest neighbors are generally simply calculated in a single space,and do not consider the correlation between multiple spaces.Based on this,we proposed a multi-label classification algorithm based on combining KNN and K-means.Firstly,KNN is used for the samples in the feature space,and the average value of the samples in each near neighbor space is taken to replace the original samples to form a new feature space.After forming a new feature space,in order to distinguish the samples clearly,K-means is used to cluster in the original feature space,and the cluster center of each sample is merged into the new feature space,the Elbow Method is used to determine the K value in K-means.The final feature space is obtained by combining the average value of samples in the neighborhood space with the new feature space.Finally,the extreme learning machine is used for classification.Experimental results show that the proposed algorithm is reasonable and effective.(2)In multi-label learning,label correlations are indispensable.It is difficult to estimate the label correlations when only part of the labels can be observed.The GLOCAL algorithm solves the problem of missing label by learning latent labels and introducing label manifold regularizes,as well as using global and local label correlations.However,when the algorithm learns the latent labels and the correlations between the original labels and the latent labels by using low-rank decomposition,the initialized low-rank matrix is randomly obtained,which may occur that the result of this algorithm is not stable.Based on this,we use K-means to cluster the original labels,and the matrix of cluster centers can correlate the original labels and the latent labels better.Experimental results show that the proposed method is reasonable and effective.