融合GCN与Informer的序列推荐算法  

A sequence recommendation algorithm integrating GCN and Informer

作　　者：范利利 李然 王宁 王客程 吴江 FAN Lili;LI Ran;WANG Ning;WANG Kecheng;WU Jiang(School of Information Engineering,Dalian Ocean University,Dalian 116023,China)

机构地区：[1]大连海洋大学信息工程学院,辽宁大连116023

出　　处：《现代电子技术》2025年第8期39-44,共6页Modern Electronics Technique

基　　金：中国医药教育协会2022重大科学攻关问题和医药技术难题重点课题(2022KTM036)。

摘　　要：为了解决长序列推荐算法的准确率低和冷启动问题,提高推荐算法的性能,提出一种融合GCN与Informer的序列推荐算法VGIN。使用图卷积网络提取数据中节点之间的空间特征,引入Informer模型来处理数据潜在的时间依赖性,再将两种特征输入多层感知器得出预测评分,实现长序列预测,改善长序列推荐效果较差的问题;同时利用变分自编码器(VAE)填补用户的数据缺失,改善用户冷启动问题。实验结果表明:构建的VGIN模型与基线模型相比得到了最高的HR@20值(0.248 4)和NDCG@20值(0.113 7),与基线版本中最优的SASRec模型相比,NDCG@20值和HR@20值分别提高了约7.87%、8.24%。该模型能有效提高长序列推荐准确率,同时降低了用户冷启动对推荐准确率的影响。In order to solve the problems of low accuracy and cold start of the long sequence recommendation algorithm and improve the performance of the recommendation algorithm,a sequence recommendation algorithm variational autoencoders graph convolutional Informer network,VGIN,is proposed.The spatial features between nodes in the data are extracted by means of graph convolutional network(GCN),and the Informer model is introduced to handle the potential temporal dependencies in the data.The two features are input into the multi-layer perceptron to obtain the scoring prediction,realize long sequence prediction,and improve the problem of poor long sequence recommendation effect.The variational autoencoder(VAE)is used to fill missing user data,thus improving the user cold start.The experimental results demonstrate that the constructed VGIN model can realize the highest HR@20 value(0.2484)and NDCG@20 value(0.1137)compared with the baseline model.In comparison with the optimal SASRec model in the baseline version,the NDCG@20 and HR@20 values can increase about by 7.87%and 8.24%,respectively.This model can effectively improve the accuracy of long sequence recommendations while reducing the impact of user cold start on recommendation accuracy.

关 键 词：序列推荐算法 冷启动 图卷积网络 Informer模型 变分自编码器 特征提取 

分 类 号：TN919-34[电子电信—通信与信息系统] TP311.1[电子电信—信息与通信工程]