Key Technique For Structure Learning Of High-Dimensional Causal Networks

Since the artificial intelligence technology is progressing gradually,the traditional research of weak artificial intelligence didn't meet researchers' requirements,and then they are more inclined to study the strong artificial intelligence technology that can solve the problem and reasoning.The research field of causality inference,mainly based on the analysis of the observational data to reveal the causal information between objects,plays a very good reasoning role in inference,prediction,decision-making,control and so on,which is one of the hot research directions in machine learning field in recent years.At the meanwhile,the telecommunications,finance,Internet,life sciences and other industries have appeared high-dimensional data research's demand along with the development of big data.Therefore there is an urgent need to solve the key and difficult points of high-dimensional causal network structure learning.In this paper,the methods of conditional independence test,Independence test and intelligent optimization algorithm are used to study,combining theory and experiment,two challenging problems:1)the reliability of independent hypothesis test based on constraint method is low,2)the overall effect of the scoring search method is not good;that need to be solved urgently in the discovery of high-dimensional causal network are deeply studied.The main contributions of this dissertation are as follows.1.A high-dimensional causal network structure learning algorithm for independent detection is studied.On the issue of causality inference,constraint-based methods are used to test the causal relationship between data,and the combinations of conditional sets gets more complex with the increasing network variables.In view of the decrease of the efficiency and accuracy of the causal network learning caused by the dimension disaster,this paper presents a new method of independent testing,which avoids the complex problem of conditional independence test.Firstly,the traditional conditional independence test method is simplified to two unconditional independence test methods to ensure the reliable test of the relationship between any two variable sets under the condition set.It is proved that the conditional independence can be derived from the two unconditional independence.The unconditional independence test is used to replace the conditional independence test on the basis of the traditional constraint-based method,and causal inference is carried out on high-dimensional data to gradually infer the complete causal network diagram.And a method presented in the paper,from the theoretical deduction and experiments,improves the accuracy of high-dimensional causal network structure learning.2.The paper studies a causal inference algorithm based on conditional set dimension reduction.In the high-dimensional data,the traditional causal network learning method has the shortcomings such as slow speed and low accuracy,and the root is complexity of the condition set of conditional independence test.In view of the above problems,firstly the mutual information method based on maximum dependence and minimum redundancy is applied to find the respective causal node set of two nodes to be tested,at the meanwhile the optimal condition set is obtained by union of two causal node sets.Using the traditional causality inference algorithm combined with the optimization condition set for causality network inference,the final inference of the complete causal network diagram comes into being.It is shown from the experimental data that the accuracy rate of the proposed method is suitable for the PC algorithm.3.A causal skeleton learning algorithm based on low-order conditional independence testing is studied.PC algorithm,especially in the network learning stage of high-dimensional data,the condition independence test has been puzzling the progress of the algorithm.Motivated by above problems,this paper in the first place puts to use the low-order condition set to carry out conditional independence testing quickly,traditional PC algorithm is utilized to calculate the rough causal network diagram,and then split the rough causal Network diagram into several small networks by the split-merge method.The network skeleton is obtained from the causal inference of the small network,and then all the small networks are combined to form a complete causal skeleton.The method presented in this paper shows that the result of this method is much better than the traditional method.4.This paper studies the causal network structure learning algorithm of high-dimensional data combined with K2 and BSO.Scoring search method also appears in the study of causal network structure,and but it is confronted with the problem that the search scoring method is unreliable under the condition of high-dimensional data.In this paper,the K2-BSO method firstly uses K2 algorithm to score the order of causal nodes such that the data population clustering process of the distance calculation function guidance algorithm gets improved.Then the population optimization is carried out by the four kinds of BSO disturbance mode until the order of causal nodes is convergent.This method settles the disadvantage that the traditional K2 algorithm is easy to get into local optimal,and improves the reliability of the algorithm in a limited time.In addition,the experimental results show that the proposed method is more reliable than the K2 algorithm and the K2-GA algorithm.Constraint based method and score based search method are two basic methods of traditional causal network learning.However,in the case of high dimensional networks,the effects of the two algorithms are unsatisfactory.Aiming at the characteristics of high-dimensional causal network and the characteristics of network learning methods,this paper proposes four algorithms for the structural learning of high-dimensional causal network,and proves the effectiveness of the algorithms theoretically and experimentally.