Research On Multi-label Classification Algorithms Based On Samples And Property Analysis

Since multi-label learning has emerged, it provides an effective solution to ambiguityproblem prevalent in various areas. It makes up for the shortcomings of traditional single-labellearning. Multi-label learning has been widely used in various research areas and become ahotspot in the areas of machine learning and data mining. The introduction of the paper brieflyreviews the research background and significance of multi-label learning and the research statusof multi-label classification algorithm in the two different directions. Then it describes in detailthe relative theoretical basis of multi-label learning, including learning framework, performanceevaluation, benchmark data sets and experimental comparison algorithms.Multi-label classification is an important part of the multi-label learning task. Whileexploring multi-label classification algorithm, most researchers focus on labels space. The mainalgorithms are broadly divided into three categories: algorithm adaptation methods, problemtransformation methods and ensemble methods. The first method improves and extends thetraditional learning algorithms to make them adapt to multi-label data. The second assigns themulti-label problems into several single label problems or regression problems. The third is theintegration of the above two methods, which combines the two methods to deal with multi-labeldata, respectively or jointly. The feature attribute space based classification algorithm takes itinto consideration that the operations on feature attribute space may influence the classificationperformance. Its effect is no less than other classification algorithms. In addition, mining thenumerical relationship between the samples and mapping to labels space provides a novel idea topredict labels.According to the multi-label classification problem, the main research works of the paperare summarized as follows:(1) Based on the multi-label classification research on feature attribute space, an improvedmulti-label scene classification algorithms based on I2C distance is proposed. Feature attribute-based algorithms which have already emerged focus on the transforming the raw data into someversion. Another method which changes feature extraction methods can also serve the purpose.In this paper, we extract the SURF feature for multi-label scene images. Instead of using a vectorto represent a sample, we adopt the vector sets to represent a sample. Secondly, we calculate thedistance between the test images and the known class by the improved I2C calculation methods.Finally, we predict all the possible labels by the correlation between labels. The results ofexperiments show that this method performs excellently on each evaluation index.(2) Based on the exploration of multi-label classification on sample numerical relationship,we propose a multi-label learning method that predicts labels for sample by mining thenumerical relationship between the samples. In traditional classification algorithms, therelationship between two samples is the similarity between them. They ignore the numericalrelationship indeed exists among samples. Exploiting sufficiently the numerical relationship among the samples and mapping to labels space in order to predict the label provide a novel ideafor multi-label classification. Firstly, the data set is divided into several groups according to thelabels. In each group, the sample matrix is processed to calculate related parameters. Then wecalculate weights of the group and neighbors by solving two optimization functions. Finally, weimprove the label prediction accuracy by controlling the weights of global mapping and localsmoothing.