| A key challenge in tumor risk prediction is selecting representative features that are responsible for tumor diagnosis.Therefore,this paper defines a Bayesian causality graph learning method based on graph constraints(BCGC)that employs causal Markov property to discover local causal dependencies between features and tumor risk types.The key point of BCGC introduces a causal network generated from an identified network skeleton to explicitly characterize these unique causal configurations of a particular tumor risk as a variable number of nodes and links.It can be analytically shown that the resulting causality graph satisfies the causal Markov property,and as a result,all local cause-effect dependencies can be retained and are globally consistent.An additional node selection estimator based on graph-constrained regression coefficients is introduced to choose the most representative features.At the same time,this paper also proposes an optimization algorithm BCGC-opt,which optimizes the order dependency between graph structure variables in BCGC.Empirical evaluations on four tumor risk datasets,and compared with traditional feature selection algorithm and causal feature selection algorithm,suggest the BCGC and BCGC-opt of this paper significantly outperforms the state-of-the-art methods.This paper contains 3 main research parts,as follows:(1)This paper introduces the origin of Bayesian causality diagrams as well as the relevant background and significance in tumor data analysis,deeply analyzes the current status of the use of Bayesian causality diagrams in the field of medical oncology at home and abroad,and summarizes and compares existing tumor diagnosis methods.Next,put forward the research content,purpose,and innovation of this thesis and the relevant theory of the Bayesian causality diagram involved in this article is described in detail.Based on the above theoretical analysis,there are several problems that need to be solved urgently: the complexity of causality graph search is high;the traditional method of selecting representative tumor features is only for the data without causality;the accuracy of tumor prediction models is low.(2)To solve the above problems and select the most representative features in tumor risk prediction,this paper proposes an improved learning method based on constraint Bayesian causality graphs —— BCGC algorithm,which combines causal Markov conditions to capture the structure of causal dependence between variables for features.Firstly,Bayesian causality graphs are used to discover the causal relationship between tumor features,and then regression coefficient estimation with graph constraints is introduced to select the most representative features.Next,the improved learning algorithm proposed in the paper is optimized.The optimization part mainly focuses on the order dependence of variables in the graph skeleton,V-structure,and orientation rules of the causal graph,so as to solve the uncertainty of the graph structure,realize parallelization and improve calculation efficiency.(3)The tumor data applied to the proposed and optimization algorithm,a majority of experiments prove that if the most representative causal feature for tumor diagnosis is selected,our method is significantly better than the existing conventional and causal features Feature selection algorithm.Besides,our method reflects model agnostics,at the same time,can reduce the complexity of the cause-effect diagram and makes model keep higher accuracy. |
