基于元学习的萌动期网络舆情预测模型

Public Opinion Prediction Model Based on Meta-Learning

现有的时间序列预测模型在网络舆情处于萌动期时,因训练所用数据大多为历史相似事件的网络舆情数据,存在严重的过拟合问题。针对上述问题,本文提出一种基于元学习的萌动期网络舆情预测模型Reptile-BiLSTM,基于SIR模型构建舆情传播网络,根据舆情特点优化OLEI算法并计算舆论的影响力特征,在历史相似事件的舆情数据上对BiLSTM模型进行预训练,使用Reptile算法在少量目标事件的舆情数据上对预训练模型进行微调,提高模型在萌动期网络舆情发展预测任务上的表现。经实验验证,Reptile-BiLSTM在evs、r2指标上较BiLSTM分别上升6%、13%,表明本文提出的模型能够在网络舆情处于萌动期时较准确地预测其发展趋势,能够尽早地为相关工作人员引导舆情发展提供有效的决策支持。

When the online public opinion is in the germination stage,the existing time series prediction model has a serious overfitting problem because most of the training data are the online public opinion data of similar historical events.To solve the above problems,this paper proposes a meta-learning based online public opinion prediction model Reptile-BiLSTM,constructs a public opinion communication network based on the SIR Model,optimizes the OLEI algorithm according to the characteristics of public opinion,calculates the influence characteristics of public opinion,and pre-trains the BiLSTM model on the public opinion data of similar historical events.The Reptile algorithm is used to fine-tune the pre-trained model on the public opinion data of a small number of target events to improve the model’s performance in predicting the development of online public opinion during the budding period.The experimental results show that the evs and r2 indexes of Reptile-BiLSTM are 6%and 13%higher than those of BiLSTM respectively,indicating that the model proposed in this paper can accurately predict the development trend of network public opinion when it is in the sprouting stage,and provide effective decision support for relevant staff to guide the development of public opinion as soon as possible.