基于混合神经网络的非法集资风险预测模型  

Illegal Fund-raising Risk Prediction Model Based on Hybrid Neural Network

作　　者：陈钢 CHEN Gang(Yangtze River Delta Information Intelligence Innovation Research Institute,Wuhu 241000)

机构地区：[1]长三角信息智能创新研究院,芜湖241000

出　　处：《计算机与数字工程》2023年第3期662-668,共7页Computer & Digital Engineering

基　　金：2021年安徽省重点研究与开发计划(编号:202104a05020071);2022年芜湖市科技计划重点研发项目(编号:2022yf17)资助。

摘　　要：针对现有企业非法集资风险识别准确率低、效率低等问题,提出了一种基于混合神经网络的预测模型。该模型构建基于预训练语言模型和门限循环神经网络(GRU)的风险等级预测网络产生风险等级和风险候选特征向量,并结合双向门限循环神经网络(BiGRU)和注意力(Attention)机制构建风险特征知识嵌入网络,最后将融合特征向量输入到分类器来实现非法集资预测。实验结果表明:该模型相较于其他基线模型能够取得更好的风险预测效果。This paper proposes an enterprises illegal fund-raising risk prediction model based on hybrid neural network to ad⁃dress the problems of the existing monitoring methods which can not identify risks effectively.The proposed risk prediction model firstly constructs a risk rank prediction network based on pre-trained language model and GRU to generate risk level and feature vec⁃tors of risk candidates.Then,a risk characteristics knowledge embedding network is created by combining BiGRU with attention mechanism to enhance risk prediction.Finally the integrated feature vectors are send to a linear classifer to finish risk prediction.Ex⁃perimental results show the proposed risk prediction model can achieve better results compared with several baseline models.

关 键 词：非法集资 混合神经网络 预训练语言模型 风险特征知识 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]