With the arrival of big data, network public opinion data reveals the features of massive information and wide coverage. For the complicated network public opinion data, traditional single models may not efficiently predict the trend of network public opinion. To address this question, the improved combination model based on the Empirical Mode Decomposition-AutoRegression (EMD-AR) model was proposed, called EMD-ARXG (Empirical Mode Decomposition-AutoRegression based on eXtreme Gradient boosting)model. EMD-ARXG model was applied to the prediction of the trend of complex network public opinion. In this model, the Empirical Mode Decomposition (EMD) algorithm was employed to decompose the time series, and then AutoRegression (AR) model was applied to fit the decomposed time series and establish sub-models. Finally, the sub-models were reconstructed and then the modelling process was completed. In addition, in the fitting process AR model, in order to reduce the fitting error, the residual error was learned by eXtreme Gradient Boosting (XGBoost), and each sub-model was iteratively updated to improve its prediction accuracy. In order to verify the prediction performance of EMD-ARXG model, the proposed model was compared with wavelet neural network model and back propagation neural network based on EMD model. The experimental results show that the EMD-ARXG model is superior to two other models in terms of the statistical indicators including Root Mean Square Error (RMSE), Mean Absolute Percentage Error (MAPE) and Theil Inequality Coefficient (TIC).