| Business changes more and more frequently due to the developing of financing. Manually analyzing and summarizing business rules by experts has been difficult to keep the speed of business changes. It is of great value to discover and valid business rules rapidly by data mining. In the background of the rule mining of reconciliation system in a large bank, this thesis systematically researched classification rule mining methods and proposed system solutions.First, this thesis introduced the background and the requirement of this project. The conversion from rule mining in finance to classification is described. Four common scenarios in financial project which are balanced and unbalanced data, labeled and unlabeled data are discussed in this thesis.Then, this thesis made a research of decision tree based classification rule mining method. For labeled and unbalanced data set, this thesis presents a classification rule mining method combined of KNN based sampling and decision tree which makes data balancing by finding the most relevant positive data samples for classification. Experiment of comparison has been made and it has been proved that rules which extracted by the classification rule mining method combined of KNN based sampling and decision tree are more concise and more in line with business logic for unbalanced data.Finally, for unlabeled balanced data set and unlabeled unbalanced data set, this thesis presents a clustering based classification rule mining method and a LOF algorithm based classification rule mining method respectively. Through the iterative process of human-computer cooperation, the workload of manually labeled data is greatly reduced under the premise of obtaining high rule accuracy. Experiment of comparison has been made and it has been proved that clustering based classification rule mining method has fast convergence rate and less data marked for balanced data, and LOF algorithm based classification rule mining method has fast convergence rate and less data marked for unbalanced data. |
