基于多信息融合分析的客户精准画像与推送算法设计  

Design of accurate customer profiling and push algorithm based on multi-information fusion analysis

作　　者：齐光鹏 QI Guangpeng(School of Physics and Electronic Engineering,Northeast Petroleum University,Daqing 163318,China;Inspur Yunzhou Industrial Internet Co.,Ltd.,Jinan 250000,China)

机构地区：[1]东北石油大学物理与电子工程学院,黑龙江大庆163318 [2]浪潮云洲工业互联网有限公司,山东济南250000

出　　处：《现代电子技术》2025年第6期175-179,共5页Modern Electronics Technique

基　　金：中华人民共和国科学技术部项目(2023YFF0905500)。

摘　　要：针对原始图卷积神经网络推送模型存在的冷启动和过平滑问题,文中基于堆叠重构网络和改进自编码器网络,提出一种针对用户画像的多信息推送模型。对于冷启动问题,在图卷积网络的输出部分,将用户画像中的评价信息嵌入到网络中,之后通过注意力网络层提取特征信息,并对模型进行堆叠,以提升用户交互数据的质量。对于过平滑问题,增加网络层数的同时,使用改进的自编码器和度预测模块对动态图网络进行局部训练,从而提升算法的个性化推荐能力。在实验测试中,相较基线最优算法,所提算法的HR指标分别提升22.7%、12.2%,NDCG指标分别提升4.7%和6.5%。证明了该算法性能良好,能够为用户提供精确化的推送服务。In allusion to the problems of cold start and oversmooth in the original graph convolutional neural network push model,a multi-information push model for user profiles is proposed based on stacked reconstruction network and improved autoencoder network.For the cold start problem,in the output part of the graph convolutional network,the evaluation information from the user profiles is embedded into the network,and then the feature information is extracted by means of the attention network layer,and the model is stacked to improve the quality of user interaction data.For the oversmooth problem,an improved autoencoder and degree prediction module are used to locally train the dynamic graph network while increasing the number of network layers,thereby enhancing the personalized recommendation ability of the algorithm.In experimental testing,in comparison with the baseline optimal algorithm,the HR index of the proposed algorithm was improved by 22.7%and 12.2%,respectively,and the NDCG index was improved by 4.7%and 6.5%respectively,proving that the algorithm has good performance and can provide users with precise push services.

关 键 词：图卷积神经网络 堆叠重构网络 用户精准画像 自注意力模型 度预测模块 推送算法 

分 类 号：TN911.73-34[电子电信—通信与信息系统] TP391[电子电信—信息与通信工程]