Research And Application Of Bayesian Network Structure Learning Algorithm Based On Inheritance

The current trend is to achieve intelligence function in various fields by uncertainknowledge inference of Bayesian networks. Bayesian network structure learning mines causalrelationship between variables and show it intuitively and comprehensively by a directedacyclic graph, which is widely concerned in both algorithm research and practical application.However, current Bayesian network structure learning algorithms can not take full accountscore decomposition and local structure inheritance when scoring, which lead to the doublecounting problem.Firstly, the individual encoding of the traditional Bayesian network structure learninggenetic algorithm needs to check the structure and ensure it acyclic repeatly. To solve thisproblem, a new individual encoding is designed, which is combined with structure inheritance.Then an improved structure learning algorithm based on inheritance and optimization isproposed and it is applied to the analysis of hypertensive disease prevention, which comes tothe conclusion that age, family history of hypertensive disorders and physique are the themain factors for hypertension. Then restrainting bad habits and maintaining healthy physiqueto prevent hypertension disease is adviced for high-risk groups.On the base of the new algorithm above, Bayesian network structure learning isresearched in the two following aspects:1) Because the BIC score is derived in large samples and does not apply to Bayesiannetwork structure learning in small samples, Weighted Score is proposed to solve this problem.And the hill-climbing algorithm is introduced for node order optimization of the K2algorithmto design the HC-K2algorithm, in which structure inheritance is taken into account.Compared to the structure learning genetic algorithm above, the HC-K2algorithm has smallersearch space of feasible solution, but it is more efficient. Then the algorithm is applied to thefailure analysis for industrial boiler in small samples, which comes to the conclusion thatwear and cracks will increase the risk of leakage and helps for boiler safety warning.2) Because multi-node Bayesian network structure learning algorithms based on nodeorder optimization of the K2algorithm are inefficient, so an improved greedy algorithmnamed Greedy-K2is proposed to solve this problem, which does not need to specify the nodeorder in advance.Compared to the HC-K2algorithm, Greedy-K2algorithm is a determinatealgorithm which comes to the only result so that it is unable to search for feasible solutions inbroader space. However, experiment shows that the algorithm is more suitable for multi-node Bayesian networks structure learning.Then the algorithm is applied to gotone customer lossanalysis for China Mobile, which comes to the conclusion that less call time last monthimplies that higher loss risk now and helps to fix the customers who are needed to be retainedThe experiment shows that the Bayesian network structure learning algorithm based oninheritance and optimization, the improved score and the Bayesian network structure learningalgorithm without specifying the node order are reasonable, effective and meaningful inpractical applications.