Multi-Label Classification With Active Learning And Semi-Supervised Bayesian Networks

The problem of multi-label is common in the real world..In these multi-labelproblems,each sample of training set is corresponding to composed of one or severalclass-label.Multi-label learning mission is forecast class-label for each unknownsample when the size of class-label set is unknown.The problem of multi-labelclassification is evidently different from traditional classification which each sampleonly belongs to one class.Single-label algorithm can’t directly used for processingthese multi-label data.Traditional multi-label classification algorithm is improved on the base ofcommon single-label algorithm.Classification results of these algorithms aresatisfactory to some extent,but all depend on the training data set need for machinelearning. The drastical cost in preparing enough labeled samples is too expensive,sohow to use large number of unlabeled samples to find the little quality samples to getthe effective classifier by training is a troublesome problem in Multi-Label field.Traditional multi-label classification algorithm requires a large number oflabeled data in order to model problem efficiently. This paper proposed a multi-labelclassification algorithm employing Gaussian fields and Bayesian network to combineactive learning and semi-supervised learning.(1)First,Multi-label data often has high dimensional feature attributes, in thiscondition, a dimension-reducing algorithm using LLE was presented to build severalBayesian network multi-label algorithms.(2)Then,on the basis of Gaussian random field model, a algorithm which isnamed ML-ASGB of combining active learning and semi-supervised of BayesianNetworks on Multi-Label Classification is proposed. This algorithm enables Bayesiannetwork multi-label classification to select a small number of high quality samplehuman label and model problem using unlabeled data.Experimental results show that M-ASGB algorithm perform better comparedwith ML_KNN and Semi-Boost algorithms, and after adding active learning thisalgorithm can decrease labeled data with acquiring equal performance compared torandom selection.