Structure Optimization Of Bayesian Network Based On High Confidence Class Label Recognition

Machine learning has long been an integral and key subfield under the domain of artificial intelligence.Among these,the problem of classification has been an important part of research in this area.Bayesian networks(BNs)have been widely used to perform knowledge representation and inference under uncertainty.It is a probabilistic graphical model with a strong probabilistic foundation,but also has the advantages of being interpretable,simple and intuitive,and of low computational complexity.The model construction process of BNC consists of two steps: structure learning and parameter learning.Structure learning involves modelling existing data and representing the implicit dependencies in the data in the form of a Directed acyclic graph(DAG),which contains both attribute nodes and class nodes,with directed edges between the nodes representing the direct dependencies that exist.Parameter learning is the learning of a conditional probability table for each node in the DAG,which is used to compute the joint probability distribution that quantifies the entire network model and obtains a set of parameters.However,for a BNC,learning the optimal BN from existing data has been shown to be NP-hard.Researchers have proposed restrictions on the network structure of BNCs to reduce construction complexity.Naive Bayes(NB)was developed to construct an extremely simplified network model by using the highly restrictive assumption of conditional independence.In the NB model,there are only directed edges from the class node to the attribute nodes,i.e.,the attributes are mutually independent of each other for a given class variable.However,this assumption is difficult to hold in complex real-life situations.This leads to a serious loss of classification accuracy for NB when dealing with complex data.Researchers have proposed various measures(e.g.attribute selection,attribute weighting,structure extension)to improve the limitations of the conditional independence assumption used by NB to some extent,resulting in a series of classical BNC models such as k-dependence Bayesian classifier,Tree-augmented naive Bayes,and Averaged Onedependence Estimators.However,different types of data sets exhibit significantly different probability distributions,and the network topology of BNC is difficult to fully represent the implied dependencies in different data sets.In addition,most BNC models assume that the instances used for training are independently and identically distributed,i.e.,if a statistical model learned from training data can achieve a correct estimate of the probability distribution,then it can also fit unknown instances in the test data.Obviously,this assumption is too strict for many applications.For BNC learning,inappropriate assumptions of independent identical distributions can lead to a single topology that does not accurately characterize the discriminative features of different classes of instances,thus negatively impacting generalization performance.In addition,traditional information-theoretic measures(e.g.mutual information)can only measure dependencies in the training data and cannot be tailored to different classification problems,ignoring information hidden in the test instances that may be beneficial for classification.The presence of low confidence class labels and their interfering information can cause misclassification and result in loss of classification accuracy.Based on the above problems,this paper proposes Class-specific Bayesian Classifier(CSB)from the perspective of asymmetric independence assertion.CSB proposes a multistage classification framework and an instance learning strategy,which can simulate the construction of arbitrary higher-order BNCs.Similar to the clustering idea,instances of the same class label should have some common implicit features,and this paper argues that they may follow the same probability distribution.Under this condition,CSB generalizes the information-theoretic-based metric functions widely used in BNC structure learning and proposes Class-MI,Class-CMI,Micro-MI,and Micro-CMI to calibrate information-theoretic measures to more finely granularly measure the dependencies between attributes(or attribute values)hidden in the training data(or test instances).In addition,CSB proposes the idea of selecting high-confidence class labels to reduce the misclassification rate and enhance the robustness of the network structure by filtering out low-confidence class labels for initial classification.CSB applies a heuristic search learning strategy to construct two sets of local BNCs corresponding to high-confidence class labels,combining them into a set of Bayesian Multinet Classifier(BMC),which can simulate the implicit dependencies between training subsets and pseudo-training subsets under the same class label,and then CSB uses integrated learning ideas for final classification.The experimental part of this paper is based on 28 UCI multiclass labelled datasets to carry out.Besides,metrics such as 0-1 loss function and root mean square error are applied to measure the classification accuracy.In addition,the Friedman test and Nemenyi test are used for the overall evaluation of the used BNC models.Experimental results show that CSB has good classification performance compared to a series of state-of-the-art single-structure models(CFWNB,AIWNB,SKDB)and ensemble models(WATAN,SA2 DE,ATODE,SLB).