摘要
针对现有的图神经网络的会话推荐模型存在引入噪声以及短期偏好片面的问题,提出一种双通道图注意力与短期偏好增强的会话推荐。通过会话序列构建会话图,并采用图神经网络提取项目的局部特征;为了过滤来自全局的噪声,设计一种双通道全局级项目表示学习层,将全局邻居节点赋予不同的权值边,区分邻居不同的转换关系达到去噪效果;设计一种短期偏好增强模块,通过注意力机制将全局信息融合,使短期偏好携带信息丰富。在Tmall、Diginetica两个公开数据集上进行实验,结果表明,所提方法优于主流模型,验证了模型的有效性和合理性。
To address the problems of noise introduction and one-sided short-term preferences in existing graph neural networkbased session recommendation models,a dual-channel graph attention network with short-term preference enhancement for session-based recommendation was proposed.A session graph was constructed from the session sequence,and local item features were extracted by using a graph neural network.To filter noise from global neighbors,a dual-channel global-level item representation learning layer was introduced,in which different weights were assigned to edges of global neighbor nodes to distinguish varying transition relationships and achieve denoising.A short-term preference enhancement module was developed to integrate global information through attention mechanisms,enriching the representation of short-term preferences.Experimental results on the two public datasets,Tmall and Diginetica,demonstrate that the proposed method outperforms mainstream models,validating its effectiveness and rationality.
作者
王一钦
吴云
WANG Yi-qin;WU Yun(State Key Laboratory of Public Big Data,Guizhou University,Guiyang 550025,China;College of Computer Science and Technology,Guizhou University,Guiyang 550025,China)
出处
《计算机工程与设计》
北大核心
2025年第12期3529-3537,共9页
Computer Engineering and Design
基金
国家自然科学基金项目(62266011)。
关键词
推荐系统
会话推荐
图神经网络
注意力机制
去噪
偏好增强
全局信息
recommendation system
session-based recommendation
graph neural network
attention mechanism
denoise
preference enhancement
global information
