Imbalanced Multi-label Learning Algorithm Based On Density Label Space

In recent years,multi-label learning has been one of the research hotspots of machine learning.With the deepening of the research,more and more problems appear.Due to the increase in the number of samples and labels,label imbalance occurs in most data sets.Some labels may have insufficient training due to too few positive category samples,thus affecting the classification accuracy.However,the existing common means to improve the imbalance is resampling and oversampling of samples,but this method has a large amount of calculation,excessive algorithm consumption,and excessive loss of sample feature information,which will lead to the decline of classification accuracy.Based on this,this thesis studies the problem of multi-label imbalance,and the specific work is as follows:1)In the multi-label data set,the label space contains a lot of hidden layer information,in which the label density information can help the algorithm effectively improve the label imbalance problem.Based on this,this thesis proposes an Imbalanced Multi-label Learning Algorithm Based on Classification Interval Enhanced,MLCIE.Firstly,the label density information is mined deeply,and the four uncertainty coefficients of each label are calculated by using these information and conditional entropy.Then the density label matrix is constructed to obtain the balanced label space.Finally,extreme learning machine is used as a linear classifier for classification.Experimental results show that the algorithm can effectively alleviate the classification errors caused by label imbalance.2)The classification performance of multi-label learning can be improved to some extent by taking care of this problem.Improving classification performance through label correlation is one of the most common and effective strategies.Many scholars have done a lot of research,but most of these studies use positive correlation-based strategies to improve performance.In practice,besides positive correlation,a negative correlation of labels may also exist.If both positive correlation and negative correlation are considered,the performance of the classifier will undoubtedly be further improved.Therefore,an imbalanced multi-label learning algorithm MLNCE is proposed in this thesis,which is based on the enhancement of negative correlation.The algorithm first uses the label density information to transform the label space,then explores the true positive and negative correlation information among labels in the density label space,and adds it to the classifier objective function.Finally,the accelerated gradient descent is used to solve the output weights to obtain the prediction results.Experimental results show that the performance of the algorithm is improved after considering both positive and negative correlations.3)Considering that the experimental data sets in the above studies are all standardized data sets,this thesis wants to study the experimental effects of previous work on real image data.Based on this,this thesis designed a multi-label natural scene image recognition application based on CNN model.In this method,CNN algorithm features are extracted from natural scene images and then MLNCE algorithm is used to classify them,thus an end-to-end recognition application is obtained.Experimental results show that this method has a certain accuracy for natural scene image recognition.Although this thesis has studied many methods to improve the multi-label imbalance,with the deepening of the research,there are still many problems of multi-label learning that have not been considered,which will be extended as the focus of work in the follow-up research.