基于EMD-XGBoost-AR模型的网络舆情预测研究

发布时间：2018-04-02 12:16

本文选题：网络舆情　切入点：网络舆情预测　出处：《广东工业大学》2017年硕士论文

【摘要】：网络舆情是网民对网络事件发表的个人观点、意见等行为倾向,当网络舆情发展到一定程度,将聚集或转化为网络舆论。习近平总书记在2016年的4.19讲话中指出“互联网已经成为舆论斗争的主战场”,网络舆论对于网络生态的正常发展起着重要性作用。若负面的网络舆论充斥着网络环境,这将影响我国的社会安定和意识形态安全。因此,及时了解网络舆情趋势,预测未来的发展态势,将有助于消解负面的舆论,维护网络生态的健康发展。由此,网络舆情的预测研究具有理论价值和社会现实意义。目前,多领域的预测研究主要采用组合预测模型用于时间序列的预测,既克服单一预测算法的缺点,更能发挥单一预测算法的优势,提高预测模型的准确性。本文基于多领域的预测模型提出了合适的网络舆情预测模型,用于提高预测模型的准确性,为网络舆情的管控提供理论支持。工作内容主要包括三个方面:(1)梳理网络舆情和网络舆情预测研究的文献内容,找出现有舆情预测的不足之处。据此引入或提出合适的网络舆情预测模型,以网民热议的“韩国部署萨德”事件和2017年“全国两会”事件作为舆情热点事件,使用微信指数、微博微指数和百度指数作为网络舆情发展趋势的量化值,进行舆情预测实验。模型预测结果以均方根误差、平均绝对百分误差和希尔不等系数作为评价标准,比较三个组合预测模型的实验结果,选取最佳的预测模型作为网络舆情预测研究的解决方案,提供新的舆情预测方法。(2)引入小波神经网络模型(WNN模型)和基于经验模态分解的BP神经网络模型(EMD-BPNN模型)用于舆情预测,并对这两个模型进行改进优化。对于WNN模型,采用附加动量法对其进行优化改进,考虑模型误差在梯度上的作用,同时还考虑误差曲面上变化趋势对模型的影响,从而避免WNN模型陷入局部最优值。对EMD-BPNN模型,采用Levenberg-Marquardt算法对其改进优化,处理模型的冗余参数问题,避免预测模型陷入局部极小值,从而加快收敛速度。(3)采用经验模态分解的算法思想,提出融合极限梯度算法(XGBoost)拟合模型的残差,在此基础上利用经典的短期时间序列模型——自回归(AR)模型用以预测网络舆情的发展趋势,即构建网络舆情预测模型EMD-XGBoost-AR模型。实验结果证明,该模型具有较好的评估性能和预估准确度。
[Abstract]:Internet public opinion is the behavior tendency of Internet users to express personal views and opinions on network events. When network public opinion develops to a certain extent, it will gather or transform into network public opinion.General Secretary Xi Jinping pointed out in his speech of 4.19 in 2016 that "the Internet has become the main battlefield of public opinion struggle" and that Internet public opinion plays an important role in the normal development of the network ecology.If the negative network public opinion fills the network environment, this will affect the social stability and ideological security of our country.Therefore, timely understanding the trend of network public opinion and predicting the future development situation will help to eliminate the negative public opinion and maintain the healthy development of the network ecology.Therefore, the prediction of network public opinion has theoretical value and social practical significance.At present, the combined prediction model is mainly used in the prediction of time series, which not only overcomes the shortcomings of the single prediction algorithm, but also improves the accuracy of the prediction model by giving full play to the advantages of the single prediction algorithm.In this paper, an appropriate network public opinion prediction model is proposed based on the multi-domain prediction model, which is used to improve the accuracy of the prediction model and provide theoretical support for the management and control of network public opinion.The work mainly includes three aspects: 1) combing the literature of network public opinion and network public opinion prediction, and finding the deficiency of public opinion prediction.Weibo micro-index and Baidu index as the quantitative value of the development trend of network public opinion, public opinion prediction experiment.The results of model prediction are based on root mean square error, mean absolute percent error and Hill unequal coefficient. The experimental results of the three combined forecasting models are compared, and the best prediction model is chosen as the solution to the research of network public opinion prediction.Provide a new method of public opinion prediction. (2) introduce wavelet neural network model (WNN) and BP neural network model based on empirical mode decomposition (EMD-BPNN) for public opinion prediction, and improve and optimize the two models.For the WNN model, the additional momentum method is adopted to optimize the model, considering the effect of model error on the gradient, and the influence of the change trend on the error surface on the model, so as to avoid the WNN model falling into the local optimum value.For EMD-BPNN model, the Levenberg-Marquardt algorithm is adopted to improve and optimize the model, to deal with the redundant parameter problem of the model, to avoid the prediction model falling into a local minimum, thus to speed up the convergence speed. (3) to adopt the idea of empirical mode decomposition (EMD).A fusion limit gradient algorithm (XG boost) is proposed to fit the residuals of the model. On this basis, the classical short term time series model, autoregressive ARL, is used to predict the trend of network public opinion, that is, the EMD-XGBoost-AR model of network public opinion prediction model is constructed.The experimental results show that the model has good evaluation performance and prediction accuracy.
【学位授予单位】：广东工业大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：G206;F49;F224

【相似文献】