UCBiG-Plugin:改进图神经网络协同过滤的通用插入式框架  

UCBiG-Plugin: A Generic Plugin Framework for Improved Collaborative Filtering of Graph Neural Networks

作　　者：潘箴烨 陈娅红 PAN Zhenye;CHEN Yahong(School of Computer Science and Technology,Zhejiang Sci-Tech University,Hangzhou 310018,China;School of Mathematics and Computer Science,Lishui University,Lishui 323000,China)

机构地区：[1]浙江理工大学计算机科学与技术学院,浙江杭州310018 [2]丽水学院数学与计算机学院,浙江丽水323000

出　　处：《软件导刊》2024年第6期1-8,共8页Software Guide

基　　金：国家自然科学基金面上项目(61772248);国家自然科学基金青年项目(11601208);浙江省自然科学基金项目(LY21A010002)。

摘　　要：图神经网络已成为协同过滤的新技术,虽然能通过迭代聚合邻域信息,自然捕获高阶的协同信号,但大部分相关工作均在在用户—物品的二部图上开展。然而,二部图中用户与物品交替连接使得用户兴趣广泛,导致在传播过程中会引入大量噪声。为此,提出一种新型的通用插入式框架(UCBiG-Plugin)直接捕获物品—物品共现图中存在的团结构,并将其粗化为新节点以构造一张全新的用户—团节点二部图,然后利用这些团结构中不同物品间存在的强接近关系,发现用户的潜在高阶语义。在3个常用的公共数据集上,应用该框架的两个改进变体进行实验评估发现,改进变体最高分别达到9.51%和8.89%,证明了同时在用户—物品二部图和用户—团节点二部图上传播协作信号能更好地捕获相关的高阶连通信息,并用于推荐任务。Graph neural networks have become a new technology for collaborative filtering.Although they can iteratively aggregate neighbor-hood information and naturally capture higher-order collaborative signals,most of the related work is carried out on the user item bipartite graph.However,the alternating connection between users and items in the bipartite graph results in a wide range of user interests,leading to the introduction of a large amount of noise during the propagation process.To this end,a new universal plug-in framework(UCBiG Plugin)is proposed to directly capture the group structures present in the item item co-occurrence graph,coarsen them into new nodes to construct a brand new user group node bipartite graph.Then,the strong proximity relationships between different items in these group structures are uti-lized to discover the potential high-order semantics of users.On three commonly used public datasets,two improved variants of the framework were applied for experimental evaluation,and it was found that the highest improved variants reached 9.51%and 8.89%,respectively.This proves that propagating collaboration signals on both user-item bipartite graphs and user-group node bipartite graphs can better capture rele-vant high-order connectivity information and be used for recommendation tasks.

关 键 词：图神经网络 团 协同过滤 推荐系统 图论 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]