多通道图神经网络的层次化融合模型用于增强会话的推荐  

HFMC-SBR:Hierarchical Fusion Model Based on Multi-channel Graph Neural Network is Used to Enhance Session-based Recommendation

作　　者：岳彩梦 彭敦陆[1] YUE Caimeng;PENG Dunlu(School of Optical Information and Computer Engineering,University of Shanghai for Science and Technology,Shanghai 200093,China)

机构地区：[1]上海理工大学光电信息与计算机工程学院,上海200093

出　　处：《小型微型计算机系统》2024年第7期1599-1607,共9页Journal of Chinese Computer Systems

基　　金：国家自然科学基金项目(61772342)资助。

摘　　要：基于会话的推荐系统(SBR)旨在根据用户历史的行为去预测下一个最有可能点击的项目.一方面由于会话推荐序列较短,可用的信息比较少,另一方面会话推荐多为匿名用户,没有丰富的用户信息,导致无法获得用户历史的交互行为或者用户的偏好,这为SBR带来了挑战.现有基于SBR研究方法大都是将会话序列建模为成对的图结构化数据或者建模为超图结构化数据,这种将会话序列建模为单一图的方法无法捕获更完整的项目转化信息,从而降低模型的准确度.为了充分考虑会话之间的相互影响,本文提出了一种多通道图神经网络的层次化融合模型用于增强会话的推荐(HFMC-SBR).模型首先将会话序列建模为全局图、局部图和超图数据,然后分别使用全局编码层和局部编码层以及超图卷机神经网络来捕获节点之间复杂的依赖性关系,学习3种项目嵌入,进而获得全局、局部以及超图项目表示信息,进而引入3层融合模型将三通道融合形成项目表示获得完整的项目转化信息,同时使用注意力机制和反向位置编码对全局上下文和局部上下文信息以及超图通道捕获的会话之间的高阶关系进行有效的融合.实验表明,本文所提出的模型HFMC-SBR,在Tmall、Diginetica和Yoochoose3种数据集上所表现的性能优于基线模型.Session-based recommender systems(SBR)aim to predict the next most likely clicked item based on the user’s historical behavior.On the one hand,the session recommendation sequence is short and the available information is relatively small;on the other hand,the session recommendation is mostly anonymous users without rich user information,resulting in the inability to obtain user history interaction behavior or user preferences,which brings challenges to SBR.Most of the existing studies of SBR only consider information of the current session,but not that of other sessions.This paper proposes a hierarchical fusion model based on multi-channel graph neural networks to enhance session recommendation(HFMC-SBR).Firstly,the session sequence is modeled as global graph,local graph and hypergraph data,and then global coding layer,local coding layer and hypergraph machine neural network are used respectively to capture the complex dependency relationship between nodes,learn three kinds of item embedment,and then obtain the representation information of global,local and hypergraph items.Furthermore,the three-layer fusion model is introduced to fuse the three-channel to form the project representation and obtain the complete project transformation information.At the same time,attention mechanism and reverse location coding are used to effectively fuse the global context information,local context information and the high-level relationship between the hypergraph channel captured sessions.Experimental results show that the proposed model,HFMC-SBR-SBR,is superior to the state-of-the-art SBR models in prediction accuracy on the three public datasets Tmall,Diginetica and Yoochoose.

关 键 词：基于会话的推荐系统 多通道信息融合 层级融合 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]