Research On Multi Label Learning Via Feature Space And Label Space Dimension Reduction Method

In multi-label learning,each instance has multiple category labels at the same time.A sample is composed of an instance and its multiple labels.The purpose of multi-label learning is to solve the ambiguity problem in text classification.Widely used in music classification,video classification,image recognition and other fields.The task of multi-label learning is to build a model based on known sample data to predict the category labels of unseen instances.With the advent of the big data era,the amount of data continues to increase sharply,and the data processed by multi-label learning usually has a large scale,which is reflected in the three aspects of the large number of samples in the data set,the higher dimension of the feature space,and the higher dimension of the label space.For multi-label learning,this situation is a huge challenge,which may lead to problems such as complex noise,high computational storage costs,and dimensional disasters.In order to overcome this challenge,a large number of multi-label learning dimensionality reduction methods have been proposed.Although such methods have achieved good performance,most of them only focus on the dimensionality reduction of a single space and ignore the high dimensions of another space.Aiming at the high dimensionality of multi-label data in feature space and label space,this paper adopts feature space and label space dimensionality reduction technology,and proposes a multi-label learning algorithm suitable for a large number of features and category labels,namely Multi Label Learning via Feature and Label Space Dimension Reduction Method(MLL-FLSDR),the main research contents are as follows:Respectively perform dimensionality reduction processing on the original feature space and label space to obtain low-dimensional feature and label space,in which the local geometric structure is maintained by the manifold constraints imposed on the low-dimensional feature space and label space,and the linear model coefficients are added.l₁paradigm constraints to filter out label-specific feature to further reduce the scale of the model,so that a multi-label classifier from low-dimensional feature space to low-dimensional label space can be constructed,and finally the decoding matrix learned in the previous stage can be used.The final prediction of the new test instance is recovered from the prediction result of the low-dimensional label space.The effectiveness of the MLL-FLSDR algorithm is verified by experiments on 17real data sets.