关键词: 神经网络模型，透明化，网络释义图，变量选择

Abstract: Since the neural network can easily fit the nonlinear mapping from the input space to the output space, the users of artificial neural network directly use it to get the black-box model with data pairs including input variables and output variables,often without taking dependencies between the input variables and output variables into account. So,there arc often redundant variables in the model which would result in poor reliability and robustness. An approach to increase visual capability for black-box properties of neural network was proposed. Firstly, the network interpretation diagram is employed to make the network transparency. I}hen, connection weights method is used to compute the relative contribution of each input variable for estimating the importance to the output variable. Lastly, the significance tests of the connection weights and contribution ratios of input variables arc implemented using the improved randomization tests for trimming the model, and the redundant variables can be eliminated by the intersection of the variables which are not significant for the overall contribution and the relative contribution rate to realize dimensionality reduction of neural network model. I}he experimental results indicate that the method can increase the transparency of model, select the best input variable set, eliminate redundant input variables, and improve the reliability and robustness of the model.Therefore, the study provides a new approach to visualize neural network model and eliminates redundant input variables.

Key words: Neural network model, Visualize,Network interpretation diagram, Variables reduction