Tax Risk Early Warning System for SMEs Using Auto EncoderBackpropagation Neural Network with Genetic Algorithm Optimization

Abstract

In the current economic circumstances, the tax risk management needs of small and medium-sized enterprises are becoming increasingly urgent. Effectively warning and controlling tax risks has become a critical issue of common concern for enterprise managers and tax departments. In view of this, the study first used the analytic hierarchy process to determine the indicator system and the corresponding weights for each indicator. Secondly, a fusion model for detection and warning was constructed by combining auto encoders with backpropagation neural networks. The performance test results showed that the optimal model parameters after genetic algorithm optimization were 8 hidden layer neurons, 11 hidden layers, dropout rate of 00.48, and learning rate of 0.024. When the amount of iterations was 1000, the loss function value of the model was 0.05, the F1 score was 0.96, the average absolute error was 0.05, and the accuracy was 0.94. In the simulation test, the model had the highest success rate of warning for the manufacturing industry, at 90.12%, and the average success rate of warning for different industries was 86.18%. The experiment findings indicated that the model exhibited high accuracy and reliability in tax risk warning. Therefore, the research not only provides a new technological means for tax risk management of small and medium-sized enterprises, but also provides a certain reference for further research and application in related fields.