首页|基于霍克斯过程和图神经网络的会话推荐

基于霍克斯过程和图神经网络的会话推荐

扫码查看
原文链接
  • NETL
  • NSTL
  • 万方数据
针对传统会话推荐系统(Session-Based Recommendation System,SBRS)往往忽略了项目点击量之间的交互,以及遗漏了会话内项目之间的相对顺序的问题,本文提出了一种基于霍克斯过程和图神经网络(Hawkes Process and Graph Neural Network,HPGNN)的会话推荐方法.该方法提出了包含图神经位置感知层和图神经霍克斯层的双流结构,分别学习用户的长期和短期偏好.图神经位置感知层通过门控图神经网络(Gated Graph Neural Network,GGNN)来捕捉各个节点之间的交互关系,得到会话中每个项目的隐向量表示,并引入逐次递减的残差网络,有效地将之前的编码信息与当前网络融合,然后通过位置感知注意力网络来捕捉项目节点在会话中的位置信息,用于学习用户的长期偏好表示.图神经霍克斯层通过将霍克斯过程和GGNN相结合来捕捉连续时间的项目点击量之间的关系,用于更准确的表示用户的短期偏好.最后将两者进行线性组合,来更好地描述用户意图.实验结果表明,提出的HPGNN在Diginetica和Yoochoose1/64两个基准会话推荐数据集上的推荐性能均优于其他会话推荐模型.
Hawkes Process and Graph Neural Network for Session-Based Recommendation
To overcome the drawbacks of traditional session-based recommendation system(SBRS)models,which of-ten ignore the interaction between item clicks and omit the relative position between items within a session,this study in-vestigated a Hawkes process and graph neural network(HPGNN)method for session-based recommendation.This novel method incorporated a dual-stream structure consisting of a graph neural position-aware layer and graph neural Hawkes layer to learn the long-and short-term user preferences,respectively.The graph neural position-aware layer cap-tured the interaction between nodes by using a gated graph neural network(GGNN)to obtain the hidden vector represen-tations of each item in a session.Furthermore,we introduced a gradually decreasing residual network that combined pre-vious encoding information with the current network.In addition,we used a position-aware attention network to capture the position information of item nodes in the session to learn the long-term preferences of users.The graph neural Hawkes layer combined the GGNN and Hawkes process to represent the users'short-term preferences more accurately by capturing the relationship between item click counts over time.Finally,we linearly combined the two layers to better describe users'intent.Experimental results showed that our proposed HPGNN outperformed other existing state-of-the-art SBRS models on the Diginetica and Yoochoose1/64 datasets.

session-based recommendationrecommender systemgraph neural networkHawkes processposition-aware attention network

杨真真、闫孟儒、杨永鹏、陈亚杰

展开 >

南京邮电大学理学院,江苏南京 210023

南京信息职业技术学院网络与通信学院,江苏南京 210023

会话推荐 推荐系统 图神经网络 霍克斯过程 位置感知注意力网络

国家自然科学基金国家自然科学基金江苏省研究生科研与实践创新计划江苏省研究生科研与实践创新计划江苏省研究生教育学改革项目南京邮电大学科研项目南京邮电大学研究生教育学改革项目

6207124262171232KYCX22_0955SJCX23_0251JGKT23_C019NY220207JGKT23_XJ02

2024

10.16798/j.issn.1003-0530.2024.04.013

信号处理
中国电子学会

信号处理

CSTPCD北大核心
影响因子：1.502
ISSN：1003-0530
年,卷(期)：2024.40(4)
  • 2