Structure Learning of CP-nets Based on Constraint and Scoring Search

CP-nets are one of the powerful tools for learning uncertain relations in the field of artificial intelligence, in which learning problems are the main research content. So far, there are many learning methods about CP-nets, which are widely used in information retrieval, user decision-making, recommendation systems and other fields. This paper attempts to propose a new solution, using an algorithm based on the combination of conditional independence and scoring search to learn the structure of CP-nets, it combines local learning, constraint-based and search scoring techniques, which is principled and effective. Firstly, this paper using the MMPC algorithm to get CPC(candidate parent-child node), and then selecting the appropriate search algorithm as a measurement standard, and perform a scoring search to obtain the optimal network structure. Based on the idea of simulated annealing in the search phase, this paper combines the probability jump feature to perform random search in the solution space to avoid falling into the local optimal solution, obtain better search results, and compare with several search algorithms. In the experimental part, it is compared with other hybrid algorithms such as sparse candidate algorithm, and the agreement is used as a measurement standard to verify the effectiveness of the algorithm.