Research On Improvement Of Neural Network Activation Function Based On Asymmetric Data

With the development and popularization of computer technology,information resources have been greatly enriched,and the problem of data explosion has followed at the same time.The original machine learning algorithm can no longer meet the needs of data mining.Therefore,under the background of big data,it is imperative to design new algorithms which can efficiently process massive amounts of data.In this paper,an artificial neural network algorithm that mimics the working principle of human brain cells is taken as the entry point,and the effect of the asymmetric sample,as well as the activation function,are proposed to improve the algorithm.Firstly,in practical applications,the asymmetry of positive and negative samples in some cases has caused great confusion in the learning algorithms.Therefore in the phase of data pre-processing before the model is established,balancing the difference between the positive and the negative samples will make the algorithm have better performance.In this paper,the principle,advantages and disadvantages of under sampling(Tomek Links etc.),oversampling(SMOTE,etc.)and comprehensive sampling method(SMOTE+ENN)are studied respectively.Finally,we chose the best SMOTE+ENN comprehensive sampling method to process the samples.As a strong expression of the non-linearity of the algorithm,the excitation function affects the accuracy of the algorithm directly.In the BP algorithm,the derivative of the activation functions also directly leads to whether the weight can converge and the speed of converge in the end.Comparing sigmoid and other activation functions,and improving the best preformed swish activation function.And then compares each parameter that changes shape of the function.Analyzing their processes in the fitting of neural network and giving suggestions of improvements.Finally,using the improvement above,we verified the data set of the default situation of credit card users in Taiwan.The comparison shows that the excitation function can change the algorithm's understanding of positive and negative class samples and achieve the requirement of improving the efficiency and accuracy of the algorithm.