融合全局信息的类别增强会话推荐方法  

Category-enhanced Session-based Recommendation Methods Fusing Global Information

作　　者：孙克雷[1] 张雪 SUN Kelei;ZHANG Xue(School of Computer Science and Engineering,Anhui University of Science and Technology,Huainan Anhui 232001)

机构地区：[1]安徽理工大学计算机科学与工程学院,安徽淮南232001

出　　处：《湖北理工学院学报》2024年第4期47-53,共7页Journal of Hubei Polytechnic University

基　　金：国家自然科学基金项目(项目编号:61703005);安徽省高校科研重点项目(项目编号:2022AH050821)。

摘　　要：现有基于会话的推荐大多只针对当前会话项目之间的顺序依赖关系进行建模,不仅忽略了其他会话中存在与当前会话相关的项目转换,而且未考虑项目类别信息在推荐过程中的辅助作用。针对这一问题,提出一种融合全局信息的类别增强会话推荐方法(CEGI-SR),学习并融合全局会话的物品转换和当前会话的物品及类别表示,使用注意力机制将当前会话中物品位置信息和交互时间信息纳入物品嵌入,辅助提升推荐效果。在两个公共数据集上进行实验,结果表明,相较最优基准模型,CEGI-SR在Diginetica数据集上P@20和MRR@20分别提高了1.15%和1.29%,在Tmall上两个指标分别提高了2.43%和2.74%,验证了CEGI-SR的有效性。Most of the existing session-based recommendations only model based on the sequential dependencies between current session items,which not only ignores the item transformations from current sessions to other sessions but also does not consider the auxiliary role of item category information in the recommendation process.To this end,this paper proposes category-enhanced session-based recommendation methods fusing global information(CEGI-SR)to improve the recommendation performance.Specifically,CEGI-SR learns and fuses the global session item transition,the current session item,and category representations.The attention mechanism is then used to include location and interaction time of the items in the current session into the item embedding to help improve the recommendation effect.The proposed CEGI-SR method demonstrates its effectiveness on two public datasets,which outperforms the optimal benchmark model with improvements of 1.15%and 1.29%in terms of P@20 and MRR@20 on the Diginetica dataset,and 2.43%and 2.74%on the Tmall dataset.

关 键 词：会话推荐 全局信息 类别信息 图神经网络 注意力机制 

分 类 号：TP391.1[自动化与计算机技术—计算机应用技术]