Research On Approximate Structure Learning Of Multi-valued CP-nets

In data mining,preference mining is an important research content.As an important expression model of conditional preference relationship,CP-nets can be used to describe the user's conditional preference relationship.At present,its main research content focuses on binary CP-nets learning,such as heuristic algorithms,evolutionary algorithms,accurate P-values,G~2 test and others.However,due to the rapid increase in the amount of data and the generation of streaming data,some learning methods of streaming data models in different periods and sections are adopted,which can reduce the model learning time and solve the online learning problem.In practical applications,binary attributes will have deviations in the expression of true preferences,the time complexity of model learning increases exponentially with the number of attributes increasing.Based on the above problems,this thesis proposes a learning method of multi-value CP-nets based on association rules and bayesian-genetic algorithm.The main research contents are as follows:(1)Research on CP-nets learning based on Apriori algorithmThis thesis proposes learning algorithm of multi-valued acyclic CP-nets based on association rules.First,according to the condition preference,the consistency scores under different parent attribute values are determined and used as the basis for judging the parent-child relationship between attributes.Then,determine the parent-child relationship candidate set is determined,and contradictory and non-existent parent-child relationships is removed with the Apriori algorithm,and the pruning operation is completed.After that,de-loop the learned CP-nets structure through the scoring function and complete the acyclic CP-nets learning.Experimental results show that this algorithm can improve the similarity and compatibility of the model effectively,and reduce the calculation time of the algorithm.(2)Research on CP-nets learning based on Bayesian-genetic algorithmAiming at the problem of multi-valued conditional preference calculation under multi-parent attribute conditions,a learning algorithm of CP-nets based on bayes-genetic method is proposed.In preference processing,the complete partial order relationship of multi-valued attributes is used as a conditional preference to determine the correlation relationship.Based on the bayesian method,a correlation database is constructed to record the correlation relationship of a single parent attribute,and to derive the correlation relationship of multi-parent attributes with a single parent attribute,reducing the calculation process.The genetic algorithm is used to search in the CP-nets structure space,and the bayesian method is used to calculate the score of a single structure to obtain the optimal structure.Through the Delink algorithm to de-loop,complete acyclic CP-nets learning.Experiments show that this algorithm effectively reduces the model learning and calculation time,and can learn local optimal acyclic CP-nets in a limited time.In summary,CP-nets learning based on the Apriori algorithm prunes through contradictions between correlation relationships,which improves computational efficiency;CP-nets learning based on Bayesian-genetic algorithm is based on Bayesian algorithm and derives the correlation relationship of multi-parent attributes through a single parent attribute,which reduces the calculation process of multi-parent attributes.The experimental comparison shows that the method in this thesis has higher similarity,compatibility and lower time complexity.Finally,the main research contents of this thesis are summarized,and the future work direction is given.