摘要
针对现有会话推荐方法主要关注用户当前会话内的短期兴趣,忽略了丰富的跨会话信息及长期兴趣信息的问题,提出了一种基于跨会话项目图的长短期兴趣推荐方法,该方法由构建跨会话项目图模块、长短期兴趣提取模块、长短期兴趣融合模块及预测模块4部分组成。该方法通过构建跨会话项目图,探索复杂的跨会话效应,采用图神经网络及多头注意力机制划分用户的长短期兴趣信息,解决偶然兴趣影响,采用门控融合机制将长短期兴趣融合为动态兴趣,预测层得到该节点的概率评分,并预测下一个点击的项目。实验在Diginetica、Yoochoose数据集上结果表明,相较于最优算法各项指标均有所提升,验证算法的有效性。
Current session-based recommendation methods primarily focus on the short-term interest of the user during a single session,ignoring the wealth of information available from cross-sessions and long-term interests.This paper proposes a long and short-term interest recommendation method based on cross-session project graph.This method consists of four parts:Cross-session Project Graph Module,long and short-term interest extraction module,long-term and short-term interest fusion module,and prediction module.This method explores the complex cross-session effects by constructing a cross-session project graph,and uses a graph neural network and a multi-head attention mechanism to the user long-term and short-term interest information,to solve the problem of occasional interest affecting,and uses a gate control fusion mechanism to fuse the long-and short-term interest into a dynamic interest,and the prediction layer gets the probability score of the node,and the next click project is predicted.Experimental results on Diginetica and Yoochoose data sets show that compared with the optimal algorithm,all the indexes are improved,which proves the effectiveness of the algorithm.
作者
李雪
周军
曲晨曦
张大俊
LI Xue;ZHOU Jun;QU Chen-xi;ZHANG Da-jun(School of Electronics and Information Engineering,Liaoning University of Technology,Jinzhou 121001,China)
出处
《计算机工程与设计》
北大核心
2025年第8期2193-2199,共7页
Computer Engineering and Design
基金
国家自然科学基金项目(12371363)
辽宁省教育厅基金项目(JYTMS20230869)。
关键词
会话推荐
跨会话
长短期兴趣
图神经网络
多头注意力机制
门控融合机制
动态兴趣
session-based recommendation
cross-session
long and short-term interests
graph neural network
multi-head attention mechanism
gated fusion mechanism
dynamic interest
