摘要
个性化推荐系统的关键是挖掘用户的情感偏好,而网络中大量的用户浏览行为记录为此提供了线索。传统描述兴趣度采用的方法是选择典型的浏览行为构造多元线性回归模型,然而浏览行为之间相互联系,容易导致共线性问题。为了提高挖掘用户兴趣的准确度,引入卷积神经网络(CNN)和胶囊网络(CapsNet),提出一种混合神经网络预测模型。首先,在卷积神经网络中对多种浏览行为使用一维卷积和最大池化操作提取局部特征;其次,在胶囊网络中将卷积网络输出的特征向量作为胶囊层的输入,使用动态路由算法对行为特征进行整体特征提取;最后,使用softmax分类器进行兴趣预测分类。实验结果表明,该模型在训练集的准确率高达95.8%,同时在测试集上的准确率都优于CNN和CapsNet,且该模型在训练过程中交叉熵损失明显低于CNN和CapsNet。采用该方法可以利用多种浏览行为准确挖掘用户的兴趣,提高了推荐系统的服务质量。
The key of the personalized recommendation system is to mine users′emotional preference.A large number of browsing behavior records provide clues to explore the users′preference.The traditional method of describing the degree of interest is to select typical browsing behaviors to construct the multi-element linear regression model.However,browsing behaviors are interrelated,which may easily lead to collinearity.In order to increase the accuracy of user interest mining,a hybrid neural network prediction model is proposed by introducing convolutional neural network(CNN)and capsule network(CapsNet).In the CNN,the local feature extraction of various browsing behaviors is implemented by one-dimensional convolution and maximum pooling.In the CapsNet,the feature vector output by CNN is used as the input of the capsule layer,and the dynamic routing algorithm is used to extract the overall feature of the behavior feature.The user interests are predicted and categorized by the Softmax classifier.The experimental results show that the accuracy of the proposed model in the training set is as high as 95.8%,and its accuracy in the test set is higher than that of the models based on CNN and CapsNet.Moreover,the model′s cross entropy loss in the training process is significantly lower than that of the models based on CNN and CapsNet.In this method,user interests can be mined accurately by various browsing behaviors,which improves the service quality of the recommendation system.
作者
王巍
洪惠君
刘阳
梁雅静
无
WANG Wei;HONG Huijun;LIU Yang;LIANG Yajing;无(School of Information&Electrical Engineering,Hebei University of Engineering,Handan 056038,China;Hebei Key Laboratory of Security&Protection Information Sensing and Processing,Handan 056038,China;School of Internet of Things Engineering,Jiangnan University,Wuxi 214122,China)
出处
《现代电子技术》
2022年第7期58-64,共7页
Modern Electronics Technique
基金
国家自然科学基金项目(61802107)
河北省物联网数据采集与处理工程技术研究中心开放课题(2016-2)
教育部-中国移动科研基金项目(MCM20170204)
江苏省博士后科研资助计划项目(1601085C)。
关键词
用户兴趣分类
混合神经网络
推荐系统
预测模型
特征提取
池化操作
user interest classification
hybrid neural network
recommendation system
prediction model
feature extraction
pooling operation
