| Summary: | With the popularity of online transactions, credit card fraud incidents are occurring more and more frequently, and adaptive enhancement (Adaboost) models are most often used in credit card fraud detection, so how to improve the robustness of the traditional Adaboost algorithm has become a hot issue. A large part of the reason for the poor robustness of the traditional Adaboost algorithm is that the base classifier is selected in a way that is uniquely oriented to the error rate. Therefore, this paper uses an adaptive hybrid weighted self-paced learning method to improve the objective function of the Adaboost algorithm, thus changing the strategy of base learner selection in the Adaboost algorithm, while the self-paced learning selected in this paper The self-adaptive threshold finding algorithm selected in this paper can well mitigate the influence of human experience on model training. This paper also selects a double-fault measure to calculate the degree of diversity among base categories from the perspective of generalization error, adds the influence coefficient of diversity to the weight calculation of weak learners, and gives the optimal range of influence coefficients through experiments. Finally, the proposed improved algorithm is applied to credit card fraud scenario, and the experiments are compared with several effective Adaboost improvement algorithms, which show that the combined performance of the proposed improved algorithm is better than other algorithms in terms of AUC value and F1 value.
|
|---|