Research On Multi-label Classification Algorithm Based On Label Relationship

Traditional machine learning algorithms generally perform modeling and prediction for a single label.In real life,a single sample often corresponds to multiple labels.Traditional machine learning algorithms cannot meet the corresponding needs and need to use multi-label learning methods to solve the corresponding problems.In most multi-label classification problems,there are usually correlations between labels.The difficulty in dealing with this type of problem lies in how to make better use of the correlation between labels for classification.There are usually two situations in the correlation between the labels: one is that the correlation between the labels is not clearly given,it is necessary to find the internal relationship between the labels through an algorithm,and then use the label correlation structure constructed by the combination of multi-label classifiers.Handle multi-label classification tasks;the other is that there is a fixed relationship between the labels.This relationship is mainly represented by the structure of a tree and the structure of a directed acyclic graph.For this situation,it is necessary to use hierarchical multi-labeling Classification model to classify the labels.Aiming at the problem of multi-label classification where the label correlation is not clearly given,this paper proposes a classifier chain(BNCC)algorithm based on Bayesian network correlation to reduce the uncertainty of the label order in the chain classifier(CC)model.The BNCC algorithm uses the neural network constructed by Tensor Flow as the classifier of all the labels,and calculates the error vector of the corresponding label to eliminate the influence of the feature set on all the labels;then,uses the error vector to identify the correlation between the labels to generate a Bayesian network,Construct a directed acyclic graph;then,determine the optimal correlation label sequence through topological sorting and reverse order;finally,use the sorted sequence as the order of the classifier chain of the CC model for classification prediction.Through experiments,the BNCC algorithm is compared with the four algorithms of CC,LIFT,BP-MLL and MLNB on five evaluation indicators and six benchmark multi-label data sets.The experimental results show that the BNCC algorithm has the better performance of dealing with general multi-label classification problems.Aiming at the problem of multi-label classification with a fixed hierarchical structure relationship between labels,this paper proposes a global classification method based on the macro-auprc optimization metric of the hierarchical multi-label deep forest(HMLDF)model,aiming at the evaluation index of hierarchical multi-label classification-macro The area under the average precision recall curve(?)adds a macro-auprc optimization method to better complete the training and prediction process of hierarchical and multi-labeled data by directional optimization of the model.The paper combines the HMLDF model and the traditional multi-label deep forest model(MLDF)based on the hloss optimization metric,as well as the improved hierarchical multi-label classification method(AWX)after the artificial neural network,are used comparative analysis was carried out in two different types of several hierarchical multi-label data sets.The experimental results show that HMLDF has better classification performance in hierarchical multi-label classification.