Research And Application Of Multi-label Learning Algorithm

With the development of computer technology and machine learning,more and more practical applications involve multi-label issues,such as image annotation,personalized recommendation and so on.An instance of multi-label problem may be associated with multiple labels or belong to multiple categories at the same time,which are significantly different from traditional classification problems.Since an instance can belong to multiple categories at the same time,the label correlations between labels becomes rich and the study of multi-label learning is more complicated.Due to the increased demand in practical applications,more and more researchers are focusing on the study of multi-label learning.Multi-label learning has become one of the hot topics in machine learning.Although great progress has been made,there are still some challenges.For example,the existing multilabel learning algorithms do not fully utilize the label correlations;in semi-supervised multilabel learning problem,where only a few training instances are available,the label correlations cannot be extended to a large number of unknown training samples;and how to apply the multi-label learning framework to practical problems.This thesis focuses on the above existing issues.First of all,the existing Calibration Label Ranking algorithm only utilizes second-order correlation information to construct the classifier,resulting that the algorithm cannot distinguish the importance of the labels during the process of label voting.To deal this issue,Calibrated Label ranking via Label Correlations Matrix algorithm is proposed.During the process of label voting,the label correlations matrix is calculated and as a weight,so that the algorithm can distinguish the importance of the labels and obtain a more reasonable voting result.Finally,a more accurate correlation label set is obtained,which effectively improves the generalization ability of the algorithm.Secondly,due to the Inductive Multi-Label Classification with Unlabeled data cannot fully utilize the second-order label correlation,an improved Semi-supervised Multi-label Learning algorithm with Second-order Label Correlation is proposed.Firstly,the binary classifiers associated with the label pairs is used as the new multi-label base classifier.Then a new calculation method for the error rate of the unknown sample is defined: adding second-order label correlation information by the combination of pairs of two label pairs to the calculation process of the error rate.As a result,the calculation of the error rate of the unknown samples is more accurate and the unknown samples' information can be used effectively.Finally,the multi-label learning is applied to the radar work mode classification.For radar work mode classification,feature parameters in each work mode are overlapping seriously.Multi-classification methods can easily cause problems such as low discrimination of each mode and low classification accuracy.To deal the problem,radar work mode classification method based on combination of multi-label learning and multi-view learning is proposed.The method has two stages: radar work mode classification based on multi-label learning and radar work mode classification based on multi-view learning of time series.In the first stage,the fuzzy theory is used to translate the radar multi-classification instances to a multilabel data set.And the training and prediction are completed with an improved calibration ranking algorithm based on the label correlation matrix.In the second stage,the work mode is classified by the historical data of the radar work mode.Finally,the two-stage classification results are merged by weighted summation to improve the classification accuracy of radar work mode.